Patrick Elliott

Patrick Elliott joined VisualCortex in April 2021 to lead the creation, development and global expansion of the company. Patrick is a veteran of the enterprise technology industry, having led teams across Canada, USA, Europe and Asia Pacific. Over the last two decades, Patrick has been predominantly based in Sydney managing teams of salespeople and solution specialists in Australia and throughout Asia Pacific. He has worked for large software companies like SAP and Oracle as well as leading teams locally for US-based companies such as Microstrategy, Cognos, Anaplan and Australian analytics success story, Yellowfin.

Q1: Tell us about how the computer vision space has evolved and your journey within it.

Patrick: Interestingly, the computer vision movement really began back in 1969 when New York City installed cameras on its municipal building. Most of the analysis was manual, but the concept of collecting data from video had begun. In fact, computer vision used to be referred to as VCA – Video Content Analysis. Today, we are actually in the fourth generation of computer vision. In essence, computer vision – or video analytics – is the ability to use machine learning models to perform automatic detections and data collection from within video footage.

Regarding modern use cases to which today’s video analytics technology can be applied, there are many possibilities – such as counting people or reading license plates. Whilst VisualCortex’s Video Intelligence Platform can facilitate any number of different computer vision use cases, internally, we stringently focus on training ML models that are globally repeatable and will provide real-world outcomes. We don’t chase every fire truck, so to speak. 

It is important to note that there are many Proof-of-Concepts in the industry or research being conducted with limited scope or depth. This is not our approach. We only work on video analytics use cases with enterprise outcomes and ML models that produce very high accuracy and confidence levels.

Q2: How does VisualCortex’s video intelligence platform connect computer vision’s potential to real-world business outcomes? 

Patrick:  VisualCortex is connecting the promise of computer vision and machine learning to real-world business cases by bringing an enterprise-ready approach to video analytics. It’s making this happen in four ways:

• Providing a lower barrier to entry by enabling organizations to use their existing infrastructure and commodity hardware
• Delivering plug-and-play access to quality-controlled models: The VisualCortex Model Store makes it easy to get started quickly with out-of-the-box, third-party, and BYO machine learning models

• Ease-of-use drives short time-to-value by enabling non-technical people to understand and act on insights derived from video analytics
• A robust platform approach, which:
  • Provides the stability to productionize machine learning models and deploy them throughout the enterprise
  • VisualCortex is also adaptable to new use cases from across different industries. It is not a point solution that’s locked into focusing on one vertical, business area, or outcome.

Q3: What are some of the biggest challenges in the video intelligence field? 

Patrick: Firstly, a lot of people are too ambitious on day one. You should focus on the most accessible use cases, which can deliver the most value, in the least time, and progress from there. From a technology standpoint, we see organizations most often struggling with the following: 

• Camera side or point solutions – which lack the scalability and flexibility to apply to meaningful, and multiple, business challenges 
• Tools that are hard to deploy commercially – financially and technologically
• Technology that’s aimed at data scientists, who aren’t business decision-makers

Q4: What are some ways that customer data can be acted on?

Patrick: Whilst we are under NDA, we can give you some good examples, which ostensibly pertain to people and vehicle detection. We’ve already completed a transport-related Proof-of-Concept with Servian, which analyzed road network utilization for a government client. This included automating things like vehicle counting, traffic pattern analysis, and anomaly detection.  

We are also working with well-known commercial real estate conglomerates to improve things like car park management. This can include things like: 

• Lowering ongoing operational costs by streamlining entry and exit protocols with automated License Plate Recognition (Our current version of LPR in car parks is yielding greater than 95% accuracy)
• Increasing revenue by improving the percentage of accurate license plate reads
• Enhancing customer experience by reducing wait times and eliminating error-based delays

In the retail space, we’re helping a nationally recognized brand to: 

• Track foot traffic effectively, efficiently, and repeatably
• Optimize in-store merchandising by analyzing shopper dwell-times
• Optimize staff rostering and reduce customer service delays;
• Capture ideal customer service examples for staff training purposes.

The more our dialogue progresses, the more use cases are emerging.

We’re also embarking on a Smart Cities initiative, which will be globally repeatable and is a natural extension of the work we do with people and transportation.

Q5: How does VisualCortex differ from its competitors? 

Patrick: Unlike camera-side or point solutions – which typically focus on one video analytics challenge per deployment – VisualCortex delivers a highly performant enterprise-grade platform to facilitate any real-time or historical video analytics use case. Because it was designed for enterprise deployment from the ground-up – rather than a result of a speculative Proof-of-Concept – it’s easy to deploy, use and add new use cases as they emerge. 

Clients are able to leverage commodity hardware to perform video analytics at scale. Many video analytics solutions rely on in-camera technology or specific types of cameras, thus forcing new hardware purchases and significantly adding to the Total Cost of Ownership. 

Through the platform approach, we can readily adapt to changing client needs. Once we’ve started a video analytics project, we often find that three or four other use cases quickly emerge. With VisualCortex, you can securely and efficiently run multiple production-level machine-learning models on each of your video sources.

VisualCortex can also be deployed anywhere clients desire – from on-premise and on-the-edge to public and private clouds, or a hybrid approach. This significantly reduces the barrier-to-entry for our customers and means they can get started faster.

Q6: How does VisualCortex store, analyze, and act on all customer data?  How does VisualCortex protect the privacy and security of its customers?

Patrick: The question of data security and privacy is one we think about a lot. Firstly, no matter where we process video content, we do not store that content.

Once we’ve processed the video in question, we discard the footage itself. What we keep is the video metadata, which is what we use to fuel use cases – to identify defined objects and actions as they occur within the original video source files. 

In terms of data processing, we can do that on-the-edge or in the cloud. The client can decide what makes sense for them from an efficiency and security perspective.

As with data processing, we have have a range of deployment options: 

1. On our cloud, as a fully-managed service; 
2. The clients’ cloud; 
3. Public cloud; 
4. On-premise and at the edge; 
5. Or hybrid

It’s completely up to the customer.

Q7: What advice do you have for organizations looking to harness video analytics technology? What critical success factors do you recommend they consider?

Patrick: There’s four core pieces of advice I can offer here: 

• Keep your project scope tight and don’t let it creep 
• Align chosen computer vision use cases to business value – how will this enhance operational efficiency, increase revenue or improve strategic decision-making?
• Establish a pragmatic rollout strategy: Tackle the most accessible and impactful use cases first to achieve early success and demonstrate value to executive sponsors
• Engage your stakeholders throughout each phase to ensure the insights being produced are as useful as hoped and are actually being used

Q8: What do you foresee as the biggest trends in computer vision and video intelligence in 2023?

Patrick: Today, the video analytics market is highly fragmented. Specific solutions are being produced by consultancies to help deliver Proof-of-Concepts. Hardware manufacturers and surveillance companies are also developing in-camera AI capabilities.

As more organizations look to harness video content for analytics and decision-making, the mushrooming number of ad-hoc video analytics solutions will become problematic. Organizations don’t want the expense and hassle of purchasing and deploying new solutions and cameras to meet each computer vision use case. 

Increasingly, organizations will need a single platform – capable of running multiple ML models across a multitude of existing video sources – to produce analytical insights about all their video content in one place. From people counting and dwell-time analysis to determine customer exposure and engagement in retail settings; to vehicle detection and tracking to better understand road usage.

The post Exclusive Talk with Patrick Elliott, CEO and CoFounder of VisualCortex appeared first on MarkTechPost.

Q1: Tell us about Klippa’s journey so far. Can you shed some light on some of the key problems that Klippa addresses?

Yeelen: Back in 2015, I was a self-employed entrepreneur, and I was working as a consultant. In both jobs, I saw how much time was spent on administrative tasks: doing your tax filings, registering your hours, approving invoices, all these kinds of tasks. I wondered why we have so many repetitive tasks that limit us from doing important things. Why couldn’t this be automated? 

I saw a lot of administrative work that nobody wanted to do. Still, you have to do it because of legal obligations: because of tax obligations or some kind of workflow within a company. We started Klippa to help companies and people automate administrative tasks so they can spend their time on more important things. 

My employer at the time was one of our first clients. They realized that if they automated some of these procedures, their consultants could spend more time with clients. Time with clients is where the money is made. Our business model is focused on saving time and reducing turnaround times in these procedures so you can spend more time doing the things that generate revenue.

Q2: How is Klippa using artificial intelligence to make intelligent document processing solutions?

Yeelen: Our space is called the IDP or intelligent document processing space. What it basically does is try to mimic the human understanding of documents or other types of files. So we use artificial intelligence to understand:

• The appearance of what a document or image looks like 
• What does it actually say? (What is on this document, and what is its meaning?) 

That combination helps us to automate certain tasks such as sorting documents, extracting certain information from documents into certain files, detecting fraudulent activity in documents, and more. We try to implement human understanding of documents and images into our solutions to allow automated decision-making.

Q3: Tell us about your IDP solution Klippa DocHorizon.

Yeelen: Klippa DocHorizon is a platform you can use as an API or an SDK. You can integrate a platform into your own workflow, your own software. Klippa DocHorizon does everything from scanning, uploading, or emailing documents to understanding that document and extracting certain information, classifying it and grouping it into certain categories, and verifying it for third-party sources. 

For example, we take an invoice to extract the Chamber of Commerce number and the VAT number. Then, we actually verify with third-party sources whether that information is valid. Does this company exist? Is it actually registered at that address? Then we convert that information into a software format that our client needs. For example, a JSON format or XML or CSV file. We turn unstructured information into structured information with all the steps in between. Most of our clients then integrate that into their software solutions.

Client using our software mainly come from the following key verticals:

• (1): Accounting and ERP industry: 
  • Companies selling bookkeeping or ERP software.
  • Use case: use it to automate receipt and invoice flows (basically Accounts Payable automation).
• (2) Loyalty and retail space:
  • Companies that collect receipts, price tags, and all kinds of information about store purchases.
  • Use case: to understand people’s behavior and in-store pricing.
• (3) KYC (know your customer) industry: 
  • Companies that need to verify the identities of their clients, such as banks, insurance companies, and car rental companies.
  • Use case: where you have to send in, for example, a passport, driving license, or an ID card to verify that you are you. 

Q4: The automation industry is seeing the rising importance of Big Data and AI. How do you see these emerging technologies impacting the Document Processing solutions sector?

Yeelen: A high volume of documents is being processed worldwide; billions of documents are probably processed daily. These are the receipts that you receive, the invoices that you receive, the contracts that you sign and the salary slips that you get. Daily, you probably touch ten documents or more, whether that’s private or within a company. 

Document processing is a field that is very applicable to using AI because there is such a large volume of available data. At the moment, probably 30% of the industry has some kind of automation around the document flows. Every year, this is growing by at least 10%.

Q5: What are the key points to consider when selecting an IDP solution?

Yeelen: When selecting an IDP solution, I think one of the core components to look at is how integratable it is. Can you integrate it with your existing solutions? Or is it a standalone solution that is not integrated? We believe that in the current day and age, any good software solution has API’s, and you can integrate one solution with the other. Almost every big company and also smaller companies have multiple software solutions. Services should be able to integrate with each other. 

For cross-border companies, it’s also very important to look at the language capabilities of IDPs. IDP solutions usually have some language limitations. Some solutions are only capable of processing in Dutch or German or French. It’s important to look at solutions that have cross-border capabilities. If you do business cross-border, which is very common in Europe and other places around the world nowadays, make sure your solution is not too limited. You want to confirm that it can support you in all your business endeavors now and in the future. 

It’s also very important to look at solutions still in development. It’s risky in software to buy a solution at some level where further developments are limited. You need to find software that is still being worked on a daily basis. You want to know that you are buying a solution that is future-proof. You’re not buying a solution that was created ten years ago. Make sure you have vendor software that has new releases every week, month, or quarter. This is very relevant so that you know that what you are buying stays up to date and is improved regularly. This is because the market is growing so fast, and the software industry is changing so fast.

Q6: Can you give our readers some tips to ensure they trust their documents to the right partner?

Yeelen: Documents can contain sensitive information, so assessing vendors’ security policy and hosting possibilities is important. Specifically in the European Union, for example, GDPR is relevant, so we need to find a partner to host your data within the European Union. But you also need to find a partner that has regular penetration testing and good internal security policies that limit their data storage. 

For example, we don’t do any data storage. We only do data processing to reduce the risks of potential data leaks. I think those are the points you should look at when looking at potential IDP partners. You should always test solutions in real life. Use real documents, test them, try them out, see the accuracy of the solutions, and then go into buying. I would never advise anyone to buy software without doing a test run. You don’t buy a car without doing a test ride at the car company. So why would you buy software without testing? 

Q7: Can you tell us about your funding status and revenue? Where are you standing right now?  

Yeelen: The software landscape has changed in the past few months due to major economic changes. It’s very nice to see now that profitable companies are still growing rapidly, and are more respected now than before. We pride ourselves on being one of those profitable companies. We are still able to grow by 100% every year without any additional funding.

Thanks to Klippa for the thought leadership/ Educational article. Klippa has supported this Content.

Please Don't Forget To Join Our ML Subreddit

The post Exclusive Talk with Yeelen Knegtering: CEO at Klippa appeared first on MarkTechPost.

Dr. Jennifer Prendki is the founder and CEO of Alectio, the first startup focused on DataPrepOps, a portmanteau term that she coined to refer to the nascent field focused on automating the optimization of a training dataset. She and her team are on a fundamental mission to help ML teams build models with less data (leading to both the reduction of ML operations costs and CO2 emissions) and have developed technology that dynamically selects and tunes a dataset that facilitates the training process of a specific ML model. 

Prior to Alectio, Jennifer was the VP of Machine Learning at Figure Eight; she also built an entire ML function from scratch at Atlassian, and led multiple Data Science projects on the Search team at Walmart Labs. She is recognized as one of the top industry experts on Data Preparation, Active Learning and ML lifecycle management, and is an accomplished speaker who enjoys addressing both technical and non-technical audiences.

Q1:Please tell us about your journey in AI so far.

Dr. Jennifer: The story of how I got into AI is a bit unusual: I actually started my career as a physicist. Understanding how the universe works was my childhood dream. But while I fulfilled that dream, the Great Recession, which started when I graduated with a Phd in Particle Physics, left few opportunities for fundamental researchers like myself to find funding, and that’s how Fate led me to pivot into a career in the industry as a data scientist instead. Still, people are often intrigued by that career shift (which to me isn’t really one since the skills I use working with data are actually so similar to the ones I used in the context of Physics research), and wonder how I ended up in the ML space from there. As far as I am concerned, there are lots of similarities between the motivations that brought me to Physics, and the reasons why I enjoy ML research. Physicists are usually motivated by a deep interest in understanding how things work – how the universe came to exist, how planets revolve around the sun or how electricity is generated -, and data scientists share this fascination for modeling the systems around them, though it is often at a more modest scale. Granted, identifying emotions from a human face on a picture isn’t as grandiose as explaining the Big Bang, but it is still extremely satisfying to figure out which exact facial features are responsible for making someone look happy, sad or angry.

Q2: Tell us about your venture Alectio and the ML technology behind it.

Dr. Jennifer: The story behind Alectio starts long before the day I incorporated the company in January 2019, and it trickles down from my frustration as a Data Science manager, as well as my legendary aversion to inefficiency in general. Back to my days at Walmart Labs, when I first led a group of ambitious data scientists, I quickly realized that my job wasn’t as much about guiding my team into choosing the right models or helping them develop new algorithms as I had originally imagined: instead, most of my time was spent begging management to grant us additional resources to get our models trained or our data annotated. And back then, every victory was short-lived, as it felt like the very moment I would receive that long anticipated approval email informing me we’d been granted a 10% increase in budget, I would also almost instantly receive another email from a team member stating that that additional budget wouldn’t cut the mustard if we wanted to meet the Black Friday deadline. I started cursing Big Data for making my life miserable, until one day, it hit me: maybe we should stop thinking of Big Data as a sine qua non condition to the success of ML projects. Even if the entire industry seemed to believe that Big Data was key to building better models, could it be that the need for large volumes of data was just a myth after all? I started researching techniques capable of reducing the amount of necessary training data and experimenting with them, and quickly concluded that working with large volumes of data was what data scientists were defaulting to because they really had no idea how to strategically sample their dataset and identify which records benefited the training process of their model. In other terms, relying on big Data was the easy thing to do back when we didn’t know any better. That’s not to say that large datasets don’t benefit ML (they absolutely do!), but they do so only because collecting more data is the path of least resistance to obtain a large enough variance, and to cover most corner cases. But in a world where Big Data was becoming a challenge (for cost or time reasons) rather than an ally, it was time to become smarter about the data we were using.

This is how I started evangelizing the concept of data curation and to promote the usage of Active Learning (a training paradigm where a model is trained incrementally with data that’s strategically selected from a raw dataset) in the industry, and promoting the idea of “Machine Teaching”, where Machine Learning is adopted to support and improve the training process of another ML model. Alectio’s technology is founded on the idea that most datasets contain a large fraction of redundant or irrelevant information which could be flushed out without impacting negatively the performance of the model, and leverages semi-supervised Machine Learning and Reinforcement Learning to establish a framework to separate useful from useless data. Today, companies rely on this technology not only to reduce the operational costs of ML development, but also to tune their data collection process and even for data explainability.

Q3: What is DataPrepOps and how is Alectio helping companies approach data collection and data preparation?

Dr. Jennifer: To better understand what DataPrepOps is, it might be a good idea to analyze how MLOps came to maturation over the past few years. Practical ML applications have flourished in the past 10 years, mostly because the hardware required to train ML models and collect sufficiently enough training data had finally caught up with the advances made in ML during the previous couple of decades. And yet, until recently, in spite of the large amount of money deployed on ML initiatives, most projects failed to launch because the people capable of building ML models were not trained to deploy them to production. Instead, it was up to DevOps teams to take on the challenge of scaling, monitoring and managing these models, so that they could actually benefit the end user and lead to ROI. And soon, as the industry started establishing expertise and best practices for model deployment and management, entire companies dedicated themselves to building tools to make that process easier and even attempt to fully automate it: MLOps was born.

Yet, MLOps fails to assist ML experts in what should be the very first step when building a Machine Learning model: constructing and optimizing a training dataset. Even in 2022, data preparation is still almost completely done by hand, which causes “data prep” to be perceived as a boring, low-tech activity, giving it an incredibly bad press among data scientists. A shocking thought when you think that preparing the data should in fact be the main center of interest of anyone working with Machine Learning. Making data the first citizen of the ML process is the very mission of the Data-Centric AI movement. DataPrepOps takes the idea one step further and addresses the reason why data prep is so unpopular with data scientists by converting it into a high-tech discipline requiring mathematical models and engineering expertise. Just like MLOps changed the game by enabling ML scientists to deploy models with no preliminary DevOps expertise, DataPrepOps leverages the most recent advances in ML to make data prep less frustrating, less expensive and overall simpler. It essentially attempts to change the perception of a field traditionally viewed as tedious and ‘uncool’, make it into a technical field in its own right, and to encourage more researchers to concentrate their efforts on technology-driven data preparation.

Q4: What are some of the biggest challenges in big data and how to solve them?

Dr. Jennifer: I believe it would be more accurate to say that it is high time we recognized that Big Data is, in itself, the challenge! The concept of Big Data was born when companies went into a frenzy to collect every single bit of data, at any cost, enabled by the fact that technology had finally made that possible for them after decades of frustration. For the longest time in the history of Machine Learning, researchers were struggling to make concrete advances precisely because the hardware available to them at the time did not allow them to collect large enough datasets to train their model on, and Big Data became the welcome antidote to the problem. What this has led to, is an exploding number of data warehouses which today are mostly filled with “ROT” (the acronym for Redundant, Outdated and Trivial) data, benefiting no one but those making money on data storage. Fundamentally, Big Data should be nothing more than a tool for data analysts and ML experts to drive decisions, build models and solve business problems. But the day the industry started treating Big Data as a field in its own right marked the beginning of what could be called the “Data Hoarding Era”, and the time when the Machine Learning field went from data-starved to data-drowned, and developing ML models suddenly became cost-prohibitive. People often don’t realize it, but Big Data actually caused ML to become more exclusive, reserved only to the few companies able to afford the preparation and storage of data at scale, as well as the associated compute costs. And that still doesn’t account for the environmental challenges posed by Big Data: with ever more data generated and collected, comes a growing need for data warehouses, and for power to supply energy to power up server farms. Without a conscious community effort to fight the “Big Data Lobby”, it is only a matter of time before only the largest corporations can even afford to train a Machine Learning model, and the advent of the super-large, extremely data-greedy models (such as GPT-4 which is coming out soon) is definitely not helping. This is why smart data curation is such an important part of the future of Machine Learning.

Q5: Multiple industries are seeing a rising importance of Big Data and AI. How do you see these emerging technologies impacting Data Privacy?

Dr. Jennifer: Anytime a company decides it’s okay for them to collect and store their user’s data  (especially without being fully transparent about it) is an issue in terms of Data Privacy, regardless of the volume collected. Sadly, both because users were only made aware of the poor practices adopted by companies worldwide years after arbitrary data collection became the norm, and because the amount of data collected was originally too low to raise the alarm, it is only recently that the world started demanding action and transparency. Too late to fully address the problem in a satisfactory way and defend the rights of consumers? Only time will tell. I am still hopeful though that as the Machine Learning field keeps maturing, new techniques based on technologies like Federated Learning (a process that allows to train on distributed, decentralized data stored at the point of collection) will allow users to maintain ownership of their data while allowing ML scientists to leverage that same data for training. At Alectio, we actually identified the protection of data ownership as a key value proposition and have successfully built on top of standard Active Learning in order to select useful data for our users while allowing them to retain their data on their systems without the need for them to export their data to the Cloud. So building Privacy-By-Design systems is certainly achievable, including in the MLOps space which usually raises concerns in that regard. And I certainly hope to see more companies follow our lead and create solutions to enable a more ethical use of users’ data in the next couple of years.

Q6: What would your advice be to budding machine learning and data science candidates? Is your company currently hiring?

Dr. Jennifer: Machine Learning is undoubtedly one of the most exciting fields in Technology nowadays, and it is only natural that so many young people are interested in building a career as data scientists. Unfortunately, after having been coined the Sexiest Job of the Century, Data Science is also attracting many people for the wrong reasons. The reality is that, just like any other career, there are some deeply frustrating aspects to a career in Data Science. People routinely believe that all data scientists are ridiculously well paid, enjoy a great work-life balance, and work on the coolest problems, but the reality is somewhat different. For example, as relatively new fields, Machine Learning and Data Science often suffer from a general misunderstanding from the people in decision making positions, which leads to many bosses having unrealistic expectations, and almost anyone in the field has a story or two to tell about how they were asked to build models with virtually no data to work with, or with no data pipelines to collect training data with. Also, not every problem out there is necessarily glamorous, and many data scientists spend their time working on detecting credit fraud or predicting when inventory is going to run out as opposed to developing DeepFake technology. My advice to aspiring ML scientists is to give themselves time to figure out if this is really the right career for them by taking internships and working on real-life ML problems. It’s also important for them to understand that “toy problems”, like the ones people work on in school or during Kaggle competitions, don’t give a reliable idea of what the life of a data scientist might be like. Being a data scientist can be cool but it’s just not for everyone, and there are many alternative careers in the data space which can be just as rewarding.

As for Alectio, we’re almost continuously hiring for ML scientists interested in a different challenge. As I often like to say, we technically are the only “true” Data Science company since our focus is on building a general framework to understand how data affects the learning process of a Machine Learning model, and how models actually learn. Besides, we’re one of very few companies using Machine Learning to facilitate Machine Learning (our ML-driven data prep algorithms are effectively ML models controlling other ML models!); in a sense, we’re in fact a Meta Machine Learning company. So if you’re really interested in fundamental ML research and in the next generation of ML algorithms, reach out to us and let’s talk!

Q7: Can you name some AI / data science resources that have influenced your thoughts the most?

Dr. Jennifer: Data scientists are spoiled with an incredible diversity of amazing content out there for them to enjoy and learn from. Regardless of your level of expertise and the specific skills you’re looking to improve, there definitely is a great Data Science blog somewhere meant just for people like you. And that is just perfect because as a data scientist, you will constantly need to stay up-to-date with the newest techniques and technology, and cannot afford to overlook the importance of continuous learning to your career. That being said, I personally have a sweet spot for the Towards Data Science and the BLAIR blogs. I also strongly recommend new-comers to the field to read the ML / AI blogs from their favorite companies, which typically provide a more specialized view on the work of ML experts within those industries and can offer more tactical / less theoretical tips for the working ML scientist.

The post Exclusive Talk With Dr. Jennifer Prendki, CEO of Alectio (An AI Startup Focused on DataPrepOps) appeared first on MarkTechPost.

Kimberly Powell is vice president of healthcare at NVIDIA. She is responsible for the company’s worldwide healthcare business, including hardware and software platforms for accelerated computing, AI, and visualization that power the ecosystem of medical imaging, life sciences, drug discovery, and healthcare analytics. Previously, Powell led the company’s higher education and research business, along with strategic evangelism programs, NVIDIA AI Labs, and the NVIDIA Inception program with over 8,500 AI startup members. Powell joined NVIDIA in 2008 with the responsibility for establishing NVIDIA GPUs as the accelerator platform for medical imaging instruments. 

Q1. Tell us about your journey at NVIDIA.  How has the advent of the NVIDIA GPU transformed the application of AI in healthcare?

Kimberly: My journey at NVIDIA started 14 years ago in the medical devices sector.  When I started, NVIDIA was primarily known for computer graphics, and over time, NVIDIA has expanded into other areas, including supercomputing and artificial intelligence.

GPU & Computer Graphics

NVIDIA’s foundational invention was the graphics processing unit (GPU).  The GPU’s purpose is a very high-level, parallel processing unit to run certain applications at orders of magnitude faster than CPUs or other architectures. The GPU is what really got me excited about joining the company when I did about 14 years ago.  This type of invention creates paradigm shifts in industries.

The first killer application of GPUs was computer graphics. In fact, our first application in healthcare was for computer graphics and radiology. Radiology is a field where we use devices to see inside the human body.  We wanted to be able to see things in more and more detail, with advanced imaging like in 3D MRI.

Accelerated Computing & Supercomputers

About fifteen years ago, NVIDIA expanded beyond computer graphics into an accelerated computing company.  GPU acceleration was paramount for the world’s supercomputers. Supercomputing is an area that we are still heavily involved in today.  NVIDIA is powering over 70% of supercomputers, which is pretty incredible.

One of the most important application areas of supercomputing centers globally is Life Sciences. One of the greatest challenges of humanity is to understand disease. At NVIDIA, we do that through very large-scale bioinformatics, molecular modeling, and simulation. This is the tip of the spear of what you could imagine industries like the pharmaceutical industry taking on. We always engage at that whole ecosystem level, starting at research, so that we can be at the bleeding edge of what our industry is going to look like in 5 to 10 years.

Domain-Specific Artificial Intelligence

Now, in 2022, the biggest and fastest-growing application area is artificial intelligence. AI is going beyond graphics in terms of what it’s doing for our company. It is the biggest technology force of the current time.  We have always firmly believed that. Now the mission statement of NVIDIA Healthcare is to bring that capability of artificial intelligence to the healthcare industry.

If you think about AI and the notion of intelligence, it means that it’s domain-specific.  There is a reason why doctors go to school and practice for decades before they are considered a specialist; because it’s very domain-specific. That is what we are doing in the healthcare industry.  We are taking these computational approaches that NVIDIA has pioneered from computer graphics to accelerated computing and artificial intelligence and putting them in the hands of the healthcare industry.

Back 14 years ago, when I started the healthcare practice for NVIDIA, we were getting all these early indicators saying that this sensor technology that was being invented needed a step function in terms of its computing power. All these improvements in the sensor technology put this huge strain on the downstream processing and the human interpretation of all of that data. Today, you will see NVIDIA inside of all of your modern medical devices, including CT, MRI, ultrasound, genomic sequencers, and microscopes.

Artificial intelligence is becoming the computational workhorse for medical device innovation. It is an area that NVIDIA is really, really focused on and we have built computing platforms to support this.

Q2. How would you describe NVIDIA’s role in contributing to the ecosystem of medical imaging, life sciences, drug discovery, and healthcare analytics?

Kimberly: Healthcare is a major industry. Those are four segments of the healthcare ecosystem that are also giant in nature. NVIDIA started in medical devices and it is still one of our core areas of contribution. The nature of medical device innovation has all of this high throughput data, and this is what is triggering the digital biology revolution. Genomics is one of the most intense data science areas ever because of these 3 billion letters that make up the story of each individual human.

NVIDIA has this unique view of being a computing platform company. We think about medical devices and allowing them to become more sophisticated in what they can sense and what they can build into their sensors.  We want to help the medical device sector innovate.  By helping the healthcare industry in creating all of this downstream data to really understand human disease, and then applying it to the challenges in each one of these industries.

Drug Discovery & Genomics

In drug discovery, you think about what steps need to be taken. We first have to identify the target, then identify what molecules might affect the behavior of that target protein, then tie that all together in ways that are completely in silico (completely done in the computer).  We use artificial intelligence, modeling, and simulation so we can reduce the amount of expensive time-consuming, error-prone experimentation that has been been used previously in early-stage drug development. You look at genomics being at the front of that pipeline, to really teach us about and help identify the genes that code for the proteins that cause our body to do certain things – good and bad. And then all the way through to these very, very large simulation problems. If you look at the drug discovery process that encapsulates it all: We go through genomics and proteins and molecular simulations, all the way through to clinical trials. NVIDIA is instrumental throughout the process- from identifying the genes and variants, all the way to early-stage in silico drug development, finishing with clinical trials.

Doing Things In Silico

How is this manifesting inside of an actual human, studying that and bringing it all back again and creating this sort of loop?  At NVIDIA we are trying to, as much as we can, put the ecosystem and processes in silico. Computer science approaches – whether it be scientific computing, artificial intelligence, or advanced visualization technique – can be applied to this data in new and sophisticated ways. This data processing is well beyond what a human, any one human, could really endure, take on, and make decisions about. So doing things in silico is really how we think about it.

Q3. Can you tell us a little about NVIDIA Clara and other healthcare tools that NVIDIA is working on? How are these tools impacting elements of the healthcare industry, such as radiology, medical imaging, and genomics?

Kimberly: NVIDIA is a whole stack computing company.  This has really helped the pharmaceutical industry understand us a bit more. Some people just call us a chip company, and we obviously find that really flawed with all the work that we are doing now.

Layer 1: GPU, chips, systems, data centers

The first layer of the stack is really around the GPU.  However, we have also moved well into full-on systems.  That first layer is about chips, systems, and whole data centers as our product.

Layer 2: Acceleration Layer

The second layer is the acceleration layer and is where the acceleration comes in. How do you take advantage of that architecture at more than a chip-level? At multiple chip levels? At multiple node levels? That acceleration layer is what really put NVIDIA on the map in the area of artificial intelligence, being able to do this deep learning at very, very large scale. A lot of the things we build at this second layer can be used in financial services, in autonomous vehicles, in the omniverse.

Layer 3: Industry Application Framework Layer

In these last 5 to 10 years we are developing the third layer of our company, and we call that the industry application framework layer. The goal of the third layer is to take that acceleration and system layer, and make it more domain-specific. 

For the healthcare domain, we call that industry application framework NVIDIA Clara. This framework was named after Clara Barton, the inventor of the American Red Cross, and we think of Clara as a platform to help people.  Our Clara platform builds upon those first two layers below it – the GPU system layer and data center layer – and leverages everything we do, as a 20,000 plus company. In this third layer of Clara, we focus on a few very specific areas.

Clara Holoscan:

Medical devices are our core. Over the last several years, we have developed our Software-Defined computational platform for medical devices called Clara Holoscan. This is where we are building actual specific systems so medical devices can do the end-to-end workloads that they need: everything from very high throughput sensor processing, to all of the AI processing they want to do in-device, and even doing visualization.  Think about an ultrasound machine: You have a sensor in and display out all on the same machine, and all of the AI and image processing that has to happen in between. This is a very typical pipeline for medical devices.

We are building a computational platform so that medical device manufacturers will not have to think about the nuts and bolts in this.  It really builds upon NVIDIA’s three core engines that we have now. (1) We invented the GPU. (2) We have been pioneering ARM-based processors. Our ARM CPU architecture is what powers all of our self-driving but now can be used for things like medical devices. (3) The third one we recently added to our family with the Mellanox acquisition is our data processing unit (DPU). You need to get data into the node at very, very high speeds. So we now have this three-engine architecture. And that’s, again, a very unique position of NVIDIA.

We want to make it much easier for the medical device community to take advantage of those three engines, and to really help them accelerate their innovation in that space. So Clara Holoscan is exactly that. The system architecture to the acceleration layer that sits on top of that to the domain-specific applications. If you look at Clara Holoscan, we have reference applications for endoscopy which is what is powering this minimally invasive surgical market. Clara Holoscan is one of the places that we are really, really excited about, and you are going to see a lot of upcoming development in that area.

We want these medical devices to become a self-driving car in a sense. What do I mean by that? We want to move into the software-as-a-service business model. Companies do not want to sell an instrument once and have to maintain it for 10 years. They want to be able to continue to innovate on AI applications and increase value upon the instrument that goes in it. That computational platform we built allows for that software-defined, medical device era to come to this field. Much of what we have learned from self-driving is completely applicable to the medical devices market.

NVIDIA Clara Discovery

Another area that we are greatly focused on is taking all of the work that we do in the supercomputing industry, and everything that we have learned from artificial intelligence, and bringing it into the drug discovery market. In the last 18 months, we announced NVIDIA Clara Discovery, our computational platform for drug discovery. Clara Discovery is all about the bleeding edge of AI and applying it to this very unique data within medical devices and biotech.  We are looking at data such as protein sequence data or smile strings that represent a molecule in the bioinformatics space and genomics data.

Transformers & Generative Models

I saw that Marktechpost followed transformers and generative models in 2021 and how applicable they are to these incredibly challenging datasets. Alphafold 2, enabled by transformer AI, allows you to essentially feed whole databases of protein sequence information so you can predict the structure of a protein.

We are pioneering generative models for molecule generation with AstraZeneca using something we call MegaMoIBART. Using transformers in generative models to go beyond the molecular databases that exist, because there are 10^60 potential molecules that we could build. Our databases are still quite small and we want to be able to explore as much as we can.  It has a lot of downstream applications in the drug discovery space.

MONAI & Computer Vision

I think most people know NVIDIA in the healthcare space on the imaging side. What we have done over the last four years is build an AI framework for medical AI. We call this framework the Medical Open Network for Artificial Intelligence, or MONAI.  MONAI is a PyTorch-based framework for deep learning in healthcare imaging.  MONAI is largely targeting a lot of the imaging applications – such as radiology, pathology, or real-time video – used in the surgical space. This week (January 19, 2022), we surpassed 200,000 downloads of this framework. It has all the domain-specific data ingestion, transformation, and model architectures used in this space. 

How do you deploy this into a clinical environment so you can validate it? We are working with a huge consortium of contributors with MONAI and building this application framework because we want to make it very, very accessible.

Computer Vision was one of the frontrunners of the application of AI.  Computer vision applications in healthcare have tremendous opportunities. Tens of thousands of algorithms will be developed to serve the radiology industry. You can use algorithms everywhere from capturing the right image to de-noising that image to then looking for anatomical structures in that image.  Doing all of the things that are repetitive that humans have to do and then presenting that information such that we can help the overstretched radiologists.  With MONAI, we are really excited about what’s happening in that space and we continue to put a lot of focus on that.

Data access & federated learning

Secondary to MONAI and just building the applications, we are also addressing one of the main challenges in healthcare around data access. A lot of computer vision and CNN (convolutional neural networks) approaches require a lot of data, a lot of labeled data. MONAI helps with that.  Healthcare data changes all the time. We want to enable a world that can better adapt to that.  All of a sudden, we are seeing lungs that have COVID pneumonia that we have never seen before. How can we create robust algorithms in real-time for that? We are doing that through a federated learning platform. 

NVIDIA FLARE

We recently announced open-source NVIDIA FLARE (Federated Learning Application Runtime Environment), which is our federated learning framework.  We worked with a consortium of 20 different hospitals and a model that was developed at Mass General. We delivered that model to 20 different hospitals so they could contribute learnings from their data but not have to contribute any data at all.  It created this really amazing multi-role model that predicts the oxygen need of a patient who had an x-ray and had some lab work done. It shows that the future of AI development will be in a federated manner.  Federated learning allows you to learn from data that is happening out on the edge but to not have to share that data that’s happening on the edge.

So that in a nutshell is a lot of what Clara is focused on.  We also have lots of efforts in NLP but, there is probably more than we can touch on.

Q4. How is NVIDIA planning to use federated learning within its healthcare division?

Kimberly: With open-source NVIDIA FLARE (Federated Learning Application Runtime Environment), we work with a lot of collaborators. At Mass General, they had this really neat model that used two different modality types: (1) electronic health record data and (2) X-ray data.  The two different modality types combined to make this prediction.  After we did this program, we actually package all of the training tools and the model itself into our NGC, which is essentially our AI software hub. We publish the model so that the world can take and build upon it. It is not an FDA-approved algorithm, but it is meant to be a tool to help the world build upon it: whether they want to learn how to do federated learning, or whether they actually want to take that model and build it into their own application framework to go through the FDA validation process. We see this as absolutely the future of model co-development.

NVIDIA is getting approached by a lot of the industry to do that co-development.  This is a very safe way to respect the privacy of data, but move the field forward and develop cutting-edge algorithms that can be heavily used. We are also enabling all of our other industries, whether it be our self-driving car industry, our financial services industry, or our retail industry. All industries have data governance challenges.  Data cannot be static. Data isn’t static. If you want AI to be able to deal with non-static data, it has to learn from non-static data. We believe federated learning is absolutely going to be what call AI 2.0. Federated learning will allow us to be able to take advantage of all the data that the world is going to continuously produce in a safe way.

Q5. Tell us about some of NVIDIA’s latest partnerships in healthcare and AI.

Cambridge-1

We have many partnerships going on. One of the ones that we are super excited about is built in the UK, called Cambridge-1. Cambridge-1 is dedicated to large-scale AI research in healthcare and the first NVIDIA built for external access. One of our collaborators, Kings College London, has developed some brain disease algorithms on it and they are actually deploying it into their clinical environment. We are also working with startups like Peptone, and others on Cambridge-1, which is the most powerful AI supercomputer in the UK.

We are also working with AstraZeneca on these transformer-based generative models for molecule generation, which we call MegaMoIBART.

Genomics & Oxford Nanopore Technologies

We are working with the genomics sequencing company, Oxford Nanopore Technologies. Stanford University’s Dr. Euan Ashley, had a dream of being able to more rapidly diagnose critical care patients through the use of genomic sequencing.  Oxford Nanopore Technologies and NVIDIA have been working together for many years.  By accelerating the whole pipeline – everything from the base calling on the sequencer all the way through to the variant calling that can decide which genetic disease you may be suffering from – we can more effectively intervene with treatment. We were able to take the world record from 14 hours down to seven and a half hours. NVIDIA is working on the accuracy and the speed of genomics with Oxford Nanopore.

Global Cancer Research

Some of the other partnerships from GTC Fall include cancer research. There is amazing work that the global cancer centers are now doing with AI. There are so many unmet needs in diseases and cancer is a big one of them. We are working with MD Anderson, St. Jude, Memorial Sloan Kettering, and German Cancer Research Center DKFZ.

Working with the startup community

In the last two years, some $40 billion in funding have flown into the drug-discovery startup community, and for good reason. The breakthroughs of alpha folds and protein structure prediction, the advancement of genomics, the fact that we can do more with AI and natural language processing. It’s a perfectly ripe time for these new companies to be established.  JP Morgan Health just finished up. There’s just partnership upon partnership of large pharma, partnering with these AI platform companies to really look for advances and acceleration. 

Q6. What do you foresee as the biggest challenges in 2022 onward for the AI and healthcare domain and where do you see NVIDIA fitting in?

Kimberly: A couple of challenges that come to mind revolve around innovation, complexity, ease of use, accessibility, and the development of specific tools that address some of the challenging data problems in healthcare. Can Clara Holoscan address the innovation problems?  Can all of Clara address the full-stack computing that is complex, but make it easy to use? Can we make it more accessible and state-of-the-art? Can we build specific tools and platforms that address some of the challenging data problems in healthcare?

(1) Reducing the complexity for the healthcare industry as computing complexity continues to grow.  In the healthcare sector, mastering the industry involves understanding the clinical problems, the workflows of the doctors, and the patients that they are trying to serve. It is very hard to do that extremely well and stay at the bleeding edge of computing approaches. With Clara Holoscan, we are working to take that complexity and make it very easy for healthcare industry professionals to remain focused on the problems.  We want to partner with them on making new computing approaches accessible to them so that:

●  Their innovation can be accelerated.

●  Their go-to-market can be much easier.

●  They can stay innovative by moving into the software-defined, software-as-a-service business model that they so desperately need and want.

How do we reduce the complexity for the healthcare industry so that they can bring these innovations to market sooner? There are all of these AI algorithms, but they have not been productized. There are several reasons for that. We are going to make sure that a computing platform isn’t the reason. You can build this application through our ubiquitous platform. You can deploy it in an instrument.  You can deploy it in a data center of the hospital. You can deploy it on any cloud. NVIDIA technology is homogeneous, and you can deploy it where your business model cares to have it.

(2) Being able to stay state of the art and making it easy to do that. NVIDIA wants to make AI accessible for research and discovery. Over in Germany, we are partnered with their cancer center DKFZ so that we can give their clinician / data scientists all the tools to ask as many questions as they wish and build all the AI application models using state-of-the-art approaches. 

With MONAI, we are helping doctors use AI to label images to really cut down their time of being the expert, by labeling highly required data. Our computing platforms enable that.

(3) Data accessibility problem

Third is the data accessibility problem. There are several ways you can skin that.

●  Federated learning is absolutely one way we can do it.  Federated learning is going to be a framework that is going to connect living breathing data to the evolution of models going forward. In the future, federated learning will enable us to develop robust models without sharing data.

●  The other is state-of-the-art approaches. What’s so novel about these transformers is you do not do it in a semi-supervised way (you do not have to have labeled data).  For healthcare, that is huge because we will never have enough labeled data. We did have some breakthrough research ourselves at NVIDIA. You can use transformers for natural language processing, you can use transformers for pre-construction prediction. We want to use state-of-the-art approaches that help us overcome some of the data challenges.

This interview was originally published in our AI in Healthcare Magazine (March 2022)

The post Exclusive Interview with Kimberly Powell, VP and General Manager of NVIDIA Healthcare appeared first on MarkTechPost.

Natalia Vassilieva is Director of Product, Machine Learning at Cerebras Systems, a computer systems company dedicated to accelerating deep learning. Her focus is machine learning and artificial intelligence, analytics, and application-driven software-hardware optimization and co-design. Prior to joining Cerebras, Natalia was a Sr. Research Manager at Hewlett Packard Labs, where she led the Software and AI group and served as the head of HP Labs Russia from 2011 until 2015. Prior to HPE, she was an Associate Professor at St. Petersburg State University in Russia and worked as a software engineer for several IT companies. Natalia holds a Ph.D. in computer science from St. Petersburg State University. 

Q1. Tell us about Cerebras Systems.

Natalia: Cerebras is an AI systems company. We’ve built a new type of computer system to greatly accelerate the training of deep neural networks (DNN) and open up new areas of research so that scientists and practitioners can do previously impossible work.

Q2. Can you tell us a little about your role as Director of Product, Machine Learning, at Cerebras?

Natalia: My role is rigorous. As a Director of Product, I sit between our engineers, customers, and the market overall. My role is to understand what kind of product we should be building, how it’s useful for our customers, and how it can potentially open doors in other markets. In practice, what that means is looking into trends in the industry overall. In terms of AI, we need to keep an eye on the latest research to understand where the field is going. Machine learning is a very fast-evolving field and many new research papers are published daily. 

Once new research is published, typically with some delay, enterprises adopt these new methods to make it easier for them to use in hardware and other applications. We are looking at state-of-the-art research, what customers want to do with that research, and what kind of applications our customers are seeking to solve. We ask, what kind of methods can be applied to help them with their task? Collecting data on the engineering organization’s requirements enables us to develop the next version of our product or software release.

Q3. Can you tell us a bit about your products, specifically those with applications in healthcare, pharma, and drug discovery?

Natalia: At Cerebras we built the world’s largest and fastest AI computer – the CS-2 system. It is a very powerful computer that enables you to train deep neural networks in hours or days vs the weeks or months it takes with legacy hardware. What we’re hearing from our customers is that when you’re working on cutting-edge research, time matters. Being able to train a model in hours or days means that researchers can test many more hypotheses that can lead to major scientific breakthroughs.

For example, we are working with pharmaceutical leader GlaxoSmithKline to use AI for drug discovery. They had a hypothesis that adding epigenomic data to their AI models would lead to more accurate and useful models. But they were previously unable to test this hypothesis because it would take too long to run on legacy hardware. They called us and we got them onto our CS-1 system. They were able to prove their hypothesis that by adding epigenomic data they could improve their models.

In regards to the pharma industry, there has been a rise in the quality of models when applied to modeling sequence data. You can think about the natural language of text as a sequence of characters or a sequence of words. People find out how to train efficient and representative models on that data in a self-supervised manner, where you don’t need any labels. You just feed all the text that you have, and it learns representations and can do some useful tasks for them. Many models have been designed to represent natural language and sequence data. The models created for language are directly applicable to modeling for biological tasks.

There is growing interest in working in domain-specific text. Being able to get insights from medical literature and to understand what kind of information can be derived from clinical reports or from any written text is important. In biotech, there are many examples of sequence data. Some examples of biological sequences include proteins, the sequence of amino acids, and DNA. If you want to model what happens in the genome, it’s a lot of modeling those sequences. 

These models are typically quite compute-intensive. High compute-intensive tasks require a heavy infrastructure footprint to be trained in a reasonable time. It is challenging to train at high scale on existing, traditional hardware. Our hardware is capable of accelerating the training of those types of models significantly. We are relevant to pharma because of our ability to process data faster with the CS-2 system.

Q4. What is the Cerebras CS-2 system? How does Cerebras use AI to drive faster drug discovery? How does the CS-2 differ from your competitors? 

Natalia: The Cerebras CS-2 is our second generation system. While our competitors are trying to connect together many weak processors, we have one giant processor that can train very large models. One of the main innovations in the CS-2 is the Wafer-Scale Engine which has 850,000 cores. This is significantly more than you can find on any CPU or GPU. It gives us the ability to significantly accelerate tasks that will require a lot of computing power.

With traditional hardware for high compute-intensive tasks, researchers are forced to cluster or connect together multiple traditional processors to be able to complete work in a reasonable time. It’s not very efficient. Instead of connecting multiple, small core count processors, we can use our single, large chip. It is easier to leverage the computing power of many cores, all packaged in a single device. The CS-2 system accelerates different compute tasks, such as the training of deep neural networks.

Q5. What are some of the biggest challenges that Cerebras is looking to address in healthcare and other verticals?

Natalia: Across all verticals, the main value proposition we offer is a powerful tool that allows domain specialists to complete their experiments much faster. We want to enable researchers to learn more quickly from the results of their experiments. 

The field of machine learning is, by and large, a field of trial and error. There is no golden book of rules on how to compose a specific model. Typically, you need to try many different things before converging on something that works for your problem. The speed of experimentation and the speed at which you can run those different trials is extremely important. 

With our hardware, we give researchers a way to make those trials take less time. We let them test many more hypotheses than they would be able to do otherwise. What we often find in practice is that researchers start with the one or two ideas they want to test. It often takes months to test a single idea in traditional environments. With Cerebras, you can test more ideas, and test them faster. 

The reality is if you don’t have the result of those experiments, it kind of slows down your imagination. If a tool can get some results in a matter of hours or days, the number of ideas that the researchers generate just explodes. Once a researcher can see what works and what doesn’t work, they can come up with 2-5 new ideas they want to test out. It fuels creativity and accelerates research significantly.

Q6. Can you tell us about Cerebras's latest partnerships in healthcare and AI? 

Natalia: We have several projects. Our partnerships with pharma companies provide a tool that enables them to create and develop new AI-driven methods. In the case of GlaxoSmithKline (GSK), we are helping them on the path to new therapeutics and new vaccines, while getting insights with the help from artificial intelligence along the way.

Another example is the collaboration with AstraZeneca. AstraZeneca has been interested in developing an internal search engine that will enable a question and answering engine. This Q&A engine will allow their researchers to find where to quickly access answers to questions about past research and past clinical trials. Another task has been building a domain-specific language model, which can help them build the question answering and machine translation engines.

Q7. How does the Cerebras platform give value to its customers?

Natalia: Typically in healthcare, we work with computational chemists, experts in biology and bioinformatics. Many of them are experts in machine learning, but almost none are experts in distributed programming. It really should be easy for them to test their ideas without knowing how the hardware works underneath, and without spending too much time thinking about they should optimize certain tasks. There is great value in running experiments much faster and making it easier for the researchers. Ease of use and fast experimentation is critical. And that is what our system brings to the table. 

I am from Russia, so let me share one more analogy from my university days. My first programming classes were taken when we were allowed just one hour on a computer. You needed to complete all your programs on a piece of paper first. You got one chance to test if your program runs right. You had to think really carefully about how you design that program, how you write that down, and then you either get it right or not and you don’t have any other chances. In many cases right now, researchers are in the same situation with these deep neural networks. When it takes you months to test your hypotheses, you know that you have only one shot, and it limits what you can do. 

Our system has essentially reduced the cost of curiosity. It enables you to not have to spend so many resources on checking whether your idea is worth pursuing. I can go ahead and test it and get more insight faster.

This interview was originally published in our AI in Healthcare Magazine (March 2022)

The post Understanding The Cerebras High-End Compute Power And Role in AI and Healthcare: Exclusive Talk with Natalia Vassilieva Director of Product, Machine Learning at Cerebras Systems appeared first on MarkTechPost.

Toby Lewis, Head of Threat Analysis
Prior to joining Darktrace, Toby spent 15 years in the UK Government’s cyber security threats response unit, including as the UK National Cyber Security Centre’s Deputy Technical Director for Incident Management. He has specialist expertise in Security Operations, having worked across Cyber Threat Intelligence, Incident Management, and Threat Hunting. He has presented at several high-profile events, including the NCSC’s flagship conference, CyberUK, the SANS CyberThreat conference, and the Cheltenham Science Festival. He was a lead contributor to the first CyberFirst Girls Competition, championing greater gender diversity in STEM and cyber security. Toby is a Certified Information Systems Security Professional (CISSP) and holds a Master’s in Engineering from the University of Bristol.

Q1: Please tell us about your role at Darktrace. What made you excited to join the Darktrace team? 

Toby: My role here at Darktrace is the Global Head of Threat Analysis. My day-to-day job involves looking at the 100 or so cybersecurity analysts we have spread from New Zealand to Singapore, the UK, and most major time zones in the US. My main role is to evaluate how we can use the Darktrace platform to work with our customers. How can we ensure that our customers get the most out of our cybersecurity expertise and support when using AI to secure their network? 

The other half of my role at Darktrace is subject matter expertise. This role involves talking to reporters like yourself or our customers who want to hear more about what Darktrace can do to help them from a cybersecurity perspective, discussing the context of current events. That part of my role was born out of a nearly 20-year career in cybersecurity. I first started in government and was one of the founding members of the National Cybersecurity Center here in the UK. It was a natural progression to continue my career at Darktrace. 

Let’s get back to the original question of what excited me about joining them. Over the last 15+ years, I’ve worked in threat intel incident response, incident management, and anything to do with security operations. A lot of that work was very reactive. We had to wait for somebody to become compromised, and we would then spend time understanding what was going on. What did the attackers do? How did they get there? From an attacker’s perspective, we could garner all this great threat intelligence, and we could then share that threat intelligence with whomever we thought needed protecting. But there always had to be a sacrificial lamb. There always had to be somebody who had to get hit, somebody who had to be compromised first so you could learn from their misfortune. 

One of the models that really excited me about Darktrace was the idea that it’s actually not fed by threat intelligence or knowledge of what attackers have done in the past. Using AI to learn about the defender’s environments, Darktrace protects against anything that doesn’t look like the defender (rather than what they think looks like an attacker). It is a powerful way of detecting things that have never been seen before, which was exciting for me. 

Q2: What are some of the biggest challenges companies face in terms of securing their organizations in 2022? How does Darktrace play a role? 

Toby: There have been several competing things happening simultaneously. On the one hand, you have the increasing use of SaaS and the cloud. On the other hand, we have got this big thing called Covid. Even as people return to the office, I don’t think we will lose hybrid work or working from home. 

Networks are no longer constrained to this tight, perimeter network that firewalls can secure. Your data is in the cloud, on some SaaS provider, or a third-party website. Fundamentally, you’re allowing your users to log in from home or over the internet. Across all these scenarios, your users can now access your data from wherever they are in the world. The trade-off is that if users can access your data from anywhere, so can an attacker. So, it becomes a question of how you would defend against that. 

How would you change your cybersecurity posture from a very traditional, barbed wire fence methodology, focusing on defending the physical perimeter, to something that anybody with internet access can now have a go at penetrating the network? The big thing that we have seen is the power of credentials. Under the old model, that perimeter, a username, and a password alone would have sufficed. Now, with widespread internet access, it’s no longer enough. An attacker can take advantage of the fact that people use weak passwords, that they reuse passwords, and that those passwords get compromised and leaked online. When users are reusing the same password across multiple sites, and it gets leaked, many other accounts could fall prey and be impacted. 

From a Darktrace perspective, recognizing that credentials have become a powerful tool for an attacker’s arsenal, we need to start thinking about how to defend the network. When somebody logs on with a username and password, how do you know they are whom they say? You have mechanisms like multi-factor authentication (MFA), but MFA isn’t a silver bullet. It’s not, “you have MFA, and therefore, all your security worries are over.” We know companies that construct MFA solutions still get targeted. We know there are weaknesses in some forms of MFA, such as SMS-based MFA, so we know that it can’t be a silver bullet. 

Using something like Darktrace’s Self-Learning AI is helpful to understand users’ behaviors so that when somebody does log on, we can determine whether that’s an expected behavior based on how we have seen them operate before. Then when they gain access to the environment and begin to move around laterally, and access services, all of those data points provide us a point of comparison with what we know that user has done in the past. That allows us to detect those unusual local events without firing on a known bad IP address or known string of texts from a malware beacon.

Q3: Can you comment on some of the cyber security breaches that took place in 2022? Ex: NVIDIA and Samsung 

Toby: The NVIDIA breach was interesting because when it first struck, it was maybe a day or two after the Russian invasion of Ukraine, and I think everyone was wondering, is this the retaliation? Is this the cyberwar that everyone has been predicting? NVIDIA is a strange place for a cyberwar to start, but was this what we expected to see? And then it transpired that it probably wasn’t anything to do with Russia at all. 

Ransomware is an incident that I spent the last four or five years focused on, more so than any other incident. It felt like yet another ransomware attack occurred when we first saw it. But as the incident evolved and more information came out about it, it was interesting to note that maybe this wasn’t a ransomware attack. Maybe, the motivation wasn’t purely about getting financial information or a financial advantage through ransom payment. We saw threats of attackers publishing hidden source code online. We noticed strange demands: “If you do not meet these demands, we will publish your source code.” We saw demands around things like removing rate-limiting for crypto mining. Attackers demanded that if companies weren’t willing to do that, they should at least open source their drivers and software so that attackers could do it themselves. 

It became one of the first attacks I have seen where it was not about trying to get a direct financial return. It was not about trying to have an ideological impact from a hacker activist perspective but getting a company to change its business practices. That said, there’s probably some financial gain further down the line in cryptomining using NVIDIA GPUs. 

Q4: Tell us about Darktrace’s self-learning AI. How does Darktrace use self-learning AI to stop cyber disruption? 

Toby: Darktrace’s approach is very different from other cybersecurity companies. Our focus is not on learning about the attacker and the methods they might use but on learning about the defender and building an understanding of normal behaviors within that organization. Self-Learning AI is constantly evolving and learns ‘on the job.’ 

We learn about our customers, including how users interact with their devices, how devices interact with each other, and what technologies different users use, for example. On a more technical level, we’re connecting data points based on packets hitting our sensors or when an API integration collects a log event. Over time, we learn behaviors and build a unique data set for each customer’s environment – understanding what is normal and what isn’t. From there, we can enforce that normal. If there is something anomalous or malicious, we can easily identify those behaviors and notify security teams in real-time.

Once something has been promoted as suspicious, we can then start applying some degree of cybersecurity context over the top. We determine that this is unusual, but this looks like an admin account, and it looks like they are trying to interact with your domain controller. Some cybersecurity context emphasizes that this anomalous event might be more worrying than just a strange random event in your environment. 

The key thing here is to focus not on the attackers but on the defenders. We build a very tight profile of what we understand about each of our customers because if there’s something alien to the environment – if the attacker is trying to get in or even an insider is trying to move around and access data they wouldn’t normally do – all that stands out from their normal behavior profile. Even if we didn’t know that they were mounting their attack from a known bad IP address, the behavior stands out compared to the other users in that environment. It’s certainly enough for us to believe that that activity is worth investigating. 

Because self-learning AI has such a deep understanding of environments and normal behaviors, it can autonomously respond when something deviates from that normal. Darktrace’s response capability, Antigena, can quarantine devices until the human team can respond. 

Q5: How does Darktrace’s solution contrast to other AI approaches? What makes Darktrace different from its competitors? 

Toby: There are two key differentiators to highlight when answering this question. The first comes back to the idea that we often throw around the word artificial intelligence (AI), but there isn’t just one way of doing it. When we look across organizations implementing AI into their technologies, it’s often an add-on – it’s not something at the core of what they do. But Darktrace has had AI at its core since its founding in 2013. 

There is a difference between supervised and unsupervised and between self-learning and pre-trained AI models. If you’re looking at a pre-trained AI model, you entirely rely on the training data and the information fed into the models before deploying it into a customer environment. Is that truly reflective of all the cyber threats that exist? Is it fully encompassing ransomware, cybercrime, nation-states’ sophisticated hackers, or the anonymous group targeting Russia currently? What happens if the attacker changes their tradecraft so radically that previous models no longer match the activity we’d expect to see? 

From Darktrace’s perspective, we recognize that attackers are incredibly diverse, broad, and too many to count. To try and build up a set of models based on attackers is an impossible task to perfect. Instead, that’s why we focus on the defenders. That’s the big difference between self-learning AI that learns the customer environment to differentiate between normal and anomalous behaviors. Pre-trained AI cannot evolve and detect unknown or never-before-seen threats.

The other aspect that makes us unique is our ability to not just detect and alert on activity but also to respond. We can apply cybersecurity context and take direct, targeted action with Darktrace’s Antigena autonomous response technology when we see suspicious or unusual activity. We can respond autonomously; this is important because we know that ransomware actors are more likely to attack when an organization’s defenses are at their weakest – such as when security teams sleep at 3 am on a Sunday. From a defender’s perspective, this means security teams don’t have to triage every alert and run a 24/7 security operations center. Antigena is already operating in the background, containing the attack as it’s happening, giving human teams time to wake up, respond, and understand what’s going on before a full eviction. Again, that approach is unique across the cybersecurity industry. 

Q6: What is Darktrace Cyber AI Research Center? What are some of the most innovative research and patents to come out of Darktrace’s research? 

Toby: A year ago, I joined an organization that now boasts around 60 patents. It’s an organization where R&D is at the core of what it’s doing. We have invested heavily in how we do research and development. We have a group of researchers predominately based in Cambridge that is genuinely at the center of AI research for its use in cybersecurity. 

Until now, this research just went to the core of our product to make our products better, but we are the only ones that really benefit from it and our customers indirectly. 

The idea we’re starting to think about is how we can share and publish this information. That’s ultimately how the research center originated. We are taking the current R&D work that we’re already doing to support our products and customers and asking: how can we share some of this with the broader community? How can we give a little bit of an insight into the work we’re doing? 

A part of that is about answering the questions of AI critics. AI critics will say AI is just a magic box doing stuff they don’t understand. But exposing our Research Center lets us show you how it works. Let us show you the research that underpins what we are doing. And again, that research has been ongoing since our founding in 2013. As we move forward into 2022 and beyond, we have been looking more and more into how we can use AI in different parts of these cybersecurity operations domains. 

It is not just about detection or the response that I have already alluded to; it is also starting to look at that Prevent model. What can we do to warn our customers about where the weak points are in their environments? How can we reassure our customers that we have complete visibility of their environment? Is there an area here that is a specific concern well before an attack occurs? Can we get our customers to start shoring up their defenses based on how we are using AI to identify weak points, hotspots, and bottlenecks in their environment? 

Q7: What are some of the use cases of Darktrace’s self-learning AI solution? Tell us about Darktrace’s latest partnerships in the tech industry. 

Toby: When Darktrace first started doing this work, it was geared toward the network level – packets, bits, and bytes flying around the network – and being able to profile that sort of activity and understand normal. As time has gone on, we have found more diverse ways of interacting and bringing in data from our customers. That data doesn’t just exist at a network layer; it exists in the cloud, SaaS, endpoints, and more. 

Some of the big pushes we have done in the last few years (and partly accelerated as our customers reacted to Covid) have been focused on how we integrate with other products. How do we bring their data to us? How do we bring their data to our AI so that we are better than the sum of individual parts?

I have worked with customers whose technology stacks are incredibly diverse, with many competing vendors. But, generally speaking, they operate in isolated silos. You have one product that might tell you one thing, then you copy and paste the results from there and put it in another tool. Then you allow it to churn, and you copy and paste it again. You’re bouncing from one tool to the next. From my perspective, some of the great things to see when I talk to our customers and our development teams is that we have successfully integrated with other major tech vendors. Ultimately, we want customers not to treat security as a siloed model.

One of the biggest partnerships we launched last year was with Microsoft. It’s been great to get a good, rich understanding of how Microsoft developed its telemetry, such as its security audit logs with Microsoft 365 and Defender. Now, we can bring all those data points into Darktrace, apply our AI on top, and provide an additional layer of assurance for customers using a Microsoft-first technology stack. It is a powerful way of augmenting an existing security stack.

Q8: What do you foresee as the biggest trends in cybersecurity in 2022? 

Toby: If you had asked me this three weeks ago, several things came to mind. The first was ransomware. I think ransomware has probably been one of the most significant topics in the last two years. There is this idea of attackers targeting a network for purely financial gain, locking out a network, stealing data, and encrypting it. 

One of those predictions is that companies choose to pay a ransom not necessarily because their files are encrypted but because businesses can’t continue normal operations. For all the hassle that an attacker must go through to encrypt files, maybe they don’t actually need to encrypt the files to have the same impact on a business. 

We start to see ransomware spread to other parts of the network estate and authentication services – so things like Active Directory servers, for example – such that attackers are not actually encrypting the data, but they are stopping your network from running. When people can’t run their companies, they are losing millions of dollars a day—suddenly, paying a ransom of $10 million as a one-off way to get your Active Directory controller back online doesn’t seem as bad. 

Another trend we have been following is the risk of the rise of the insider threat — this idea of the Great Resignation. People are now free to make life changes they previously held back on due to COVID lockdowns. Do they opt for a career change or move to a new job? Is this the opportunity to stay working from home as their de facto way of working? Or is their company mandating a return to the office?

We are seeing this almighty churn start with staff moving from organization to organization. Not only are they potentially a risk to their employer by walking out with sensitive data in their possession, but what does the process look like when they actually leave? How is that company locking down that environment access so that somebody can come in after that employee has left and keep that account active, logging on, and getting access to the network? 

Finally, we have the Russia-Ukraine conflict. Many assumed this would be the first major military conflict where cyber was a critical factor in deciding elements of the battlespace. Now, arguably, we have not seen that evolve as we thought. But does that still change the cybersecurity landscape? Has the invasion of Ukraine brought more Western entities together and criminal entities together to fight a common foe? Will we see a truce on the horizon? I don’t know. But certainly, it means there is a lot more uncertainty due to the global disruption. 

The post Exclusive Talk with Toby Lewis, Global Head of Threat Analysis at Darktrace appeared first on MarkTechPost.

Ryan McDonald is the Chief Scientist at ASAPP. He is responsible for setting the direction of the research and data science groups in order to achieve ASAPP’s vision to augment human activity positively through the advancement of AI. The group is currently focused on advancing the field of task-oriented dialog in real-world situations like customer care. In such dynamic environments, there are many interacting parts: the conversation between customer and agent; the environment and tools the agent is using; different measures of success; a wide range of customer needs and situations. Optimizing this environment in order to lead to quality outcomes for customers, agents and companies require significant research investment in retrieval, language generation, constrained optimization, learning, and, critically, evaluation.

Ryan has been working on language understanding and machine learning for over 20 years. His Ph.D. work at the University of Pennsylvania focused on novel machine learning methods for structured prediction in NLP, most notably information extraction and syntactic analysis. At Penn, his research was instrumental in growing the fields of dependency parsing and domain adaptation in the NLP community. After his Ph.D., Ryan joined Google’s Research group. There he researched sentiment analysis and summarization models for consumer reviews, which resulted in one of the first large-scale consumer summarization systems consumed by millions of users every day.

Q1. Tell us about your journey in AI.

Ryan: I’ve been working on language understanding and machine learning for over 20 years. My Ph.D. work at the University of Pennsylvania focused on novel machine learning methods for structured prediction in NLP, specifically information extraction and syntactic analysis. After my Ph.D., I joined Google’s Research group focused on sentiment analysis and summarization models for consumer reviews, which resulted in one of the first large-scale consumer summarization systems consumed by millions of users every day. While there, my team was instrumental in the development of Google Assistant as a global technology by building out many multilingual capabilities. After over a decade working on consumer products, I then shifted gears towards enterprise and led numerous NLP and ML projects to improve Google’s Cloud services, including the core NLP API, solutions for Call Center AI, and knowledge discovery from the scientific literature. My research on enterprise NLP and ML now continues at ASAPP.

Q2: Tell us about your role as Chief Scientist at ASAPP and your leadership style and philosophy

Ryan: At ASAPP, I’m tasked with setting the research agenda to realize our vision of augmenting human activity positively through the advancement of AI.

My experience has been that industrial research labs that are successful in the long term are those that have a culture of execution excellence to drive business objectives. As such, I am a strong believer that strong industrial AI research works backward from present or future business outcomes in order to develop a program that can be broken down into a series of short-term objectives, each of these testable. So one of my main jobs is to ensure that we start efforts by thinking about what are the outcomes we care about: are these useful? How large is the impact? Etc. From there we focus on measurement: what data do we have or need to collect? How hard will it be to get that data? What metrics can we measure and optimize that are correlated with the outcomes we care about? Only then should we think about AI models and solutions.

I have a lot of great teammates at ASAPP who are capable of amazing things. This dictates my leadership style, which is more focused on making sure the top-level objectives are aligned with ASAPP’s short/medium and long-term goals. Once that is done, I mainly focus on ensuring we have the resources in place to execute as well as working to remove obstacles.

Q3: In terms of some of the biggest challenges in the customer service and call center space, How can ML, and NLP help to improve the customer support/ agent experience?

Ryan: The customer experience and contact center industry often finds itself in a tricky balance between lowering costs while attempting to improve the quality of customer service. 

Before widestream adoption of AI, companies used a variety of methods to lower costs but suffered a lower quality of customer care in the process. “Containment,” a measure of having customers solve their own issues without human intervention, was seen as a key way to lower costs. This often came in the form of simple rule-based systems such as an interactive voice response (IVR) or chatbot which used FAQs to help customers solve their issues. And, to reduce customer wait times, new agents were given abbreviated training periods. Unfortunately, the confluence of these efforts created scenarios where customers never had their issues resolved through self-service means, and agents encountered high turnover rates from lack of training and support from automation. 

Today, almost every stage of your interaction with a call center could be driven by AI or already have AI informing how the issue is addressed. After a customer connects with an agent AI can guide and make suggestions to the agent. What should they say next? What flow should they follow? What knowledge base articles will help solve the problem? While companies should be optimizing AI in these ways, what we are finding is that most still do not. Such models are best trained on historical data and optimized for some key performance indicators, which can handle time (how quickly the problem was solved) or customer satisfaction score (was the customer happy with the experience). 

Once the call or chat is over, AI is still at work. In most call centers the agent will leave structured information and notes about what happened during the call. This is for analytics purposes but also for any subsequent agent who picks up the issue if it has not been resolved. AI helps with all these steps.

Finally, there are supervisors who are there to help assist agents and grow their skills. AI can be critical here. In a call center with hundreds of agents handling thousands of calls a day. How can supervisors identify the issues that need their intervention? How can they understand what happened during the day? How can they find areas of improvement for agents in order to grow their skillset?

At ASAPP, we’ve also found that while real-time dynamic guidance for agents is critical, more structured training, coaching and feedback is also important. Many agents train on new issues or procedures ‘live’. That is, they get a description of the procedure, but then only see it in practice when they take a call with a real customer. Imagine we gave pilots the manual of the plane and then told them to fly 300 passengers to Denver? Because of this, we are focusing on using AI to help build tools for agents to practice procedures and handle difficult situations before they deal with live customers. When this is coupled with targeted feedback (either by a supervisor or automatically) this will allow the agent to grow their skills in a less stressful environment.

Q4: Can you share a little about ASAPP’s AI Services and Platform? 

Ryan: ASAPP’s AI services integrate into existing CX environments and support people in being their best by predicting what an agent can say and do throughout every customer interaction, automate tasks within their workflow and continuously retrain the models to ensure increased accuracy and ultimately impact. 

Our AI platform has a particular emphasis on empowering agents through a host of modular AI services. Ready via API, SDK, or plug-in options, we offer the following services:

• JourneyInsight – Analyzes agent activity in depth, identifies ways to streamline
• AutoCompose – Crafts quality agent responses for digital messaging
• AutoTranscribe – Delivers highly accurate speech-to-text transcription
• AutoSummary – Creates high-quality disposition notes automatically
• CoachingInsight – Provides real-time visibility, tools to guide agent performance
• AutoWorkflow – Automates time consuming tasks for agents during interactions

Each of these is designed and trained to optimize for key business outcomes. Specifically, these services focus on the joint objective of improving the experience of the call center customer and the job satisfaction of the agent. ASAPP customers also derive greater value when multiple services are together. The network effect of using multiple AI services makes every one of them better for you. 

Q5: How is ASAPP bringing its AI technology in a way that is different from its competitors?

Ryan: Our central hypothesis at ASAPP is that AI should augment people in positive and productive ways. This takes shape in our research and product strategy which has a particular focus on the agent and their experience. We combine our domain expertise to create AI models tailored for the contact center and customer experience use-cases. 

The AI-driven results speak for themselves. An airline customer saw agent productivity increase 86% and a rise of organizational throughput (total number of interactions across all customer service channels divided by labor spent to satisfy those needs) by 127%. For a global network operator, their net promoter scores (NPS) (the willingness of customers to recommend a company’s products or services to others) increased 45%. For a telecommunications company, their cost per interaction decreased 52%. These examples show how AI, designed for people, can increase productivity, improve the quality of customer service, and decrease business costs.

Q6: Can you shed some light on the latest employment trends related to ML and NLP? Is your company hiring?

Ryan: Yes, we’re hiring! We can’t comment on behalf of all ML and NLP jobs, but at ASAPP, prioritized areas of ML research and engineering are in speech, ML engineering, and task-oriented dialog. 

Researchers at ASAPP work to fundamentally advance the science of NLP and ML toward our goal of deploying domain-specific real-world AI solutions, and to apply those advances to our products. They leverage the massive amounts of data generated by our products, and our ability to deploy AI features into real-world use to ask and address fundamental research questions in novel ways.

The post Exclusive Talk with Ryan McDonald: Chief Scientist at ASAPP appeared first on MarkTechPost.

Naveed Ahmed Janvekar is a Senior Data Scientist working at Amazon in the United States. He works on solving fraud and abuse problems on the platform that impacts millions of customers of Amazon in the US and other parts of the world using Machine Learning and deep learning. He has 7+ years of expertise in the Machine Learning space which includes classification algorithms, clustering algorithms, graph modeling, BERT to name a few. He is using Machine Learning and Deep Learning to solve multi-faceted problems. He has a Master’s degree in Information Science from The University of Texas at Dallas where he graduated top of his class and was awarded as a scholar of high distinction and inducted in the prestigious International Honor Society Beta Gamma Sigma. He has a Bachelor’s of Engineering in Electronics and Communications from India. He has worked with other influential firms such as Fidelity Investments and KPMG. In his current role, he is researching identifying novel fraud and abuse vectors on ECommerce platforms and using Active Learning to improve Machine Learning model performance.

Editor’s note: The opinions expressed in this article are solely those of Mr. Naveed Ahmed Janvekar and do not express the views or opinions of his employer.

Q1: Tell us about your journey in AI and data science so far.  What factors influenced your decision to pursue a master’s degree and a career in the field of AI?

Naveed: I am currently working as a Senior Data Scientist at Amazon, working on improving customers’ shopping experience by the detection and prevention of abusive or policy-violating entities within the platform. My journey in AI/ML started slightly before I enrolled in the Master’s program at UT Dallas. While I was working with Fidelity Investments in India, I was inspired by a couple of analysts who were making use of data to make impactful business decisions. This experience, along with my ambition for pursuing higher education, led me to study information science with a specialization in machine learning. After graduating from UT Dallas, I worked with KPMG as a Business Intelligence Developer working on building reporting applications. In 2017, I joined Amazon as a Business Analyst and worked my way upwards to become a Senior Data Scientist.

Q2: Tell us about your current role?

Naveed: My current role as a Senior Data Scientist involves building strategic Data Science roadmap/projects to continually improve customers’ shopping experience. On a day-to-day basis, I engage with various business stakeholders on various business problems and peer scientists to discuss the latest ML methodologies. Model building, experimentation, data extraction, and coding are pretty much part of the daily routine. Innovating on behalf of the customers is on the daily.

Q3: What are some of the biggest challenges as a Data Scientist?

Naveed: I believe one significant challenge is to get the right kind of data for exploration, model training, and/or insight generation. Many times, the data that is available might not be structured or even available in relational databases. There could be data quality issues, missing data, and features needed for model training might not be readily available.  In addition, being able to engineer these features can be pretty time-consuming and complex. 

Another challenge with respect to supervised machine learning models is the lack of availability of high-quality training datasets.  By high-quality, I mean aspects such as enough volume of labels, data quality, and class balance to name a few. Sometimes building a narrative and effective storytelling of any analysis or data science solution to stakeholders can be challenging based on the complexity of insights. This is something one gets better at with time and constant engagements with business partners.

Q4: What is your opinion of machine learning in the field of fraud and abuse prevention?

Naveed: Machine learning helps in the scalability and accuracy of fraud and abuse detection in a cost-effective manner. For example: if one were to manually evaluate every transaction or entity as fraud or not then there is a pretty good chance of catching all bad transactions or entities. But in today’s world the scale of transactions and interactions is huge, billions of transactions, millions of entities make it humanly not cost-effective and possible to evaluate all of them manually.  By using machine learning, one can automate fraud detection as much as possible with high predictive power and cheaper costs.

Q5: What would your advice be to budding machine learning and data science candidates?

Naveed: Data Science is considered as a generalist role by many.  Hence while having data science breadth knowledge is important, my advice would be to also to focus on data science depth knowledge such as mathematical details behind algorithms. This will give you an edge over others in the field. Also, being good at storytelling, communication insights to business stakeholders, and building a narrative around your solutions. Participate in as many data science competitions as possible, participate in publishing research papers and get mentors early on in your career.

Q6: Can you name some books, courses, or other resources that have influenced your thoughts the most? 

Naveed: My biggest influencer has been Andrew Ng, an American Computer Scientist. He has courses on Coursera which are pretty good. Hands-On Machine Learning with Scikit-Learn by O’Reilly Media is a good book as well. For programming skill improvement my go-to is leetcode for both python and SQL. The Data Incubator is good for fellowships and certifications.

Q7: What are your views about Marktechpost.com?

Naveed: I came across Marktechpost in 2021 and have been following it ever since. I really like the kind of articles and content that is being put out by this publication. Most of the latest ML innovations are covered and very nicely summarized for quick reads. This clubbed with the courses made available by Marktechpost makes it a great place for AI professionals to regularly engage with the platform. 

The post Exclusive Talk with Naveed Ahmed Janvekar: Senior Data Scientist at Amazon appeared first on MarkTechPost.

The demand for artificial intelligence (AI) is increasing rapidly. Helping young people get educated and exposed to careers in AI has become a critical need. However, there are several challenges in getting underserved communities involved in AI. Learning to develop AI applications requires access to hands-on learning and adequate computing resources. In addition, many underrepresented communities, including women and people of color, do not see many role models in the field nor do they have guidance on getting started.

NVIDIA & BCGWPA Partnership Aims to Be Part of the Solution: Making AI Education More Accessible

To address this need, NVIDIA and the Boys & Girls Clubs of Western Pennsylvania (BGCWPA) entered into a three-year partnership in April 2021. Their joint goal is expanding access to AI education and scaling the curriculum to young people in traditionally underserved and underrepresented communities nationwide through the development of the AI Pathways Toolkit and expansion of the Artificial Intelligence Pathways Institute Program to Boys & Girls Clubs outside of Western PA.

The Marktechpost team interviewed leaders from both NVIDIA and BGCWPA on their partnership and the goal for their AI Pathways Toolkit. Liz Austin, a leader of NVIDIA’s philanthropic arm, and Dr. Lisa Abel-Palmieri, CEO at BGCWPA, were interviewed.

1. Marktechpost: Can you tell us about the AI education partnership between the BGCWPA and NVIDIA?

● Dr. Lisa Abel-Palmieri, BGCWPA: Boys & Girls Clubs of Western Pennsylvania was awarded a PA Smart Grant with the whole goal of advancing STEM education in the State of Pennsylvania to ensure that we have a workforce that’s ready for the jobs of the future. And so, we initially designed this concept of the Artificial Intelligence Pathways Institute (AIPI) as part of that grant initiative. Our first cohort, our pilot program, in 2019 was about 40 young people. The majority of these young people were girls and students of color. Through local contacts, we were able to get connected with Liz and her wonderful team at NVIDIA to really just dream up and flesh out what it could look like to actually build this AI Pathways Toolkit and scale these AI knowledge & skills across the country.

● Liz Austin, NVIDIA: On the NVIDIA side, the Foundation learned about some of the early work that Lisa described with the pilot summer programs. We believe that all students should have access to AI education. We were excited about the opportunity to partner with BGCWPA to help them scale the reach of their program to more students.

2. Marktechpost: Can you please shed some light on the NVIDIA Jetson Nano 2GB developer kit and JetBot platform?

● Dr. Lisa Abel-Palmieri: The primary hardware we are working with, in terms of young people getting exposure to AI, is the Jetson Nano 2GB dev kit and open-source JetBot AI robotics platform from NVIDIA. In the first AIPI cohort, we had to figure out how the young people could use the Jetson Nano hardware to build an autonomous moving robot. Our second cohort in 2020 was a camp, and we had about 60 students. At this time, the Jetson had come further along in its development. There was an open-source curriculum we could access online that we then incorporated more formally into our AIPI program. This summer we are hosting over 200 students.

● Liz Austin: NVIDIA’s goal with the Jetson Nano 2GB Developer Kit and the new grant program is to further democratize AI and robotics. With easy-to-follow tutorials and ready-to-build open-source projects created by an active community, the Jetson Nano 2GB is ideal for hands-on learning, building, and teaching AI and robotics.  

● According to a student of the 2019 AIPI cohort, Paige Frank: “Learning robotics hands-on with the Jetson Nano made it much easier. And it was exciting to actually see our programming in action as the NVIDIA JetBot robot navigated the maze we created for the Project.” Paige is now an intern with the BGCWPA, where she mentors and helps other young teens getting started with AI and coding.

3. Marktechpost: Can you please tell us about the AI Pathways Toolkit and AI Pathways Institute?

● Liz Austin, NVIDIA: With this particular initiative, our focus is partnering with the Boys & Girls Clubs to build out this toolkit consisting of that open-source curriculum and the staff tools and training. Our goal is to make it really easy for the Boys & Girls Clubs’ staff and other educators to deliver and implement the curriculum with their students. 

● Dr. Lisa Abel-Palmieri: The hardware piece of this is key, and it’s at the heart of the Artificial Intelligence Pathways Institute. We explore other topics in our curriculum, like AI ethics and human-centered design, and how to design a product. And these additional career readiness skills, those sort of communication skills or business skills, project management, a taste of all those career readiness skills.

● Dr. Lisa Abel-Palmieri: The AI curriculum, as of now, is definitely more geared to high school students. As we build out this curriculum even further, we’re looking to create different modules. With NVIDIA foundation support, we can bring in a full-time manager whose major support and effort will be to refine and break down these modules, to be able to customize it and scale it down to appropriate audiences based on the needs.

● AIPI aims to serve 20,000+ high school-aged students in the Boys & Girls Clubs community by 2024. And that’s just the beginning! The open-source curriculum will also be available to other organizations interested in implementing AI education programs worldwide.

4. Marktechpost: How are educators involved in the AI Pathways Toolkit curriculum?

● Dr. Lisa Abel-Palmieri: I am excited about this partnership because we can scaffold the learning in a way with somebody who has no experience. The youth development professionals who work with students at Boys & Girls Clubs that use the toolkit will eventually get a certification from Jetson. It is a piece of hardware that somebody from a non-STEM background (like education or psychology) can pick up. We can walk them through how they can eventually gain certification, and that’s why we chose Jetson in the first place.

5. Marktechpost: What were some of the biggest challenges in building the curriculum?

● Dr. Lisa Abel-Palmieri: Overcoming the perception that the field is difficult, designing it such that it’s for everyone, the diversity piece. In many of these roles, whether IT or software development, there is a lack of representation, particularly among people of color and women. We hope that in making the curriculum and the hardware accessible, we will ultimately be able to get more diverse folks into the field. To really roll kids in, you get them ultimately connected by personalizing learning and creating projects where they have a voice. It’s about creating opportunities where the topics are relevant to them.

● Intentional aspects of this program will be videos and interviews and stories of diverse people talking about their career journey and how they got into the roles, these technical roles, they’re in. “You can’t be what you can’t see.”

● Students also got to visit companies and see AI in action and present a capstone project that focused on a social problem they wanted to solve with AI.

6. Marktechpost: What has the students’ experience been? Can you give us some insights on the outcome after the first cohort? What is the biggest impact you have seen in people’s lives? What have been some of your students’ most rewarding experiences?

● Dr. Lisa Abel-Palmieri: It’s a work-based learning experience where we pay all the young people a fair stipend. People are getting paid to learn, and then all of a sudden, they want to go to college for computer science. And it’s something that pays a good salary too. It makes a big difference when you can enter into these careers and break your family’s cycle of poverty. It’s literally a life-changing opportunity to enter into the technology space. Upon entering the program, only 50% of the young people were interested in going to a four-year college, and post-program surveying, we are now up to 90%. This program has shown them the possibilities and pathways that they never thought are possibilities for themselves. This program doesn’t fix college affordability, and many others are working on it as well.

● What are the students saying? Paige Frank of the 2019 pilot program says she wants to strengthen her coding skills and become a Python pro. And she definitely wants to pursue computer science in college. “I have a lot of goals,” says Paige.

7. Marktechpost: How do you envision this program growing in the future? How many students are you hoping to serve?

● Dr. Lisa Abel-Palmieri: We eventually want to be in many Boys & Girls Clubs all over this country. The goal in the next few years is that we will be running this summer intensive program to up to 80 Boys & Girls Clubs. One of our primary goals is to increase diversity and democratize AI.

● Liz Austin: Our goal is to serve 20,000 kids in Boys & Girls Clubs across the USA by 2024. And from there, the sky’s the limit, as we also plan to make the toolkit available to other organizations around the world that are interested in implementing AI education programs. 

Source:

NVIDIA: https://blogs.nvidia.com/blog/2021/04/20/ai-pathways-boys-girls-clubs/

Jetson Nano Toolkit:

https://www.nvidia.com/en-us/autonomous-machines/embedded-systems/jetson-nano/education-projects/

NVIDIA Leadership: Liz Austin – Liz Austin Author Page | The Official NVIDIA Blog

Boys & Girls Club Leadership: Dr. Lisa Abel-Abel-Palmieri | Boys & Girls Clubs of Western PA (bgcwpa.org)

The post An Interview with Key Leaders at NVIDIA and Boys & Girls Clubs of Western Pennsylvania: Their Partnership Aims to Make Artificial Intelligence More Accessible to Youth in Underserved Communities appeared first on MarkTechPost.

Asif: Tell us about your journey in AI and machine learning so far.  What factors influenced your decision to pursue a PhD and a career in the field of AI? 

Sherin: I was initially intrigued by the field of Machine Learning (ML) and Deep Learning (DL) as it presented a world of endless possibilities with applications in complex domains such as video/image classification, tracking face recognition, and biomedical signal processing. I felt there was more for me to learn, discover something new in this area, solve challenging tasks, and be able to achieve something significant. My natural curiosity and an opportunity to challenge myself were the driving forces that pushed me to explore and pursue a Ph.D. in this field.  In my Ph.D., I developed a novel dictionary and DL algorithms for classification tasks related to remote health monitoring systems (e.g., activity recognition for wearable sensors). Completing the Ph.D. degree does require years of hard work, but I think this was one of the best decisions I made in my life. This journey has enriched me, not only in terms of knowledge in the field, but also taught me how to handle setbacks and be persuasive during challenges which are important to be successful in the industry. In my industry work experience, I have developed the ability to create new ML models to improve and increase the effectiveness of cybersecurity products and work on leading edge research such as eXplainable AI (XAI) and deepfakes. I feel truly indebted to the learnings that I have received over the years and feel deeply passionate about XAI, ethical AI, the opportunity to combat deepfakes and digital misinformation, and topics related to ML and DL with applications for cybersecurity.

Asif: How does computer vision differ from human vision?  What are some of the factors you take into consideration developing machine learning algorithms?

Sherin: Computer vision (CV) allows computers to “see” via pixels and interpret digitally, incorporating pattern recognition and mimicking human vision by recognizing objects in images or videos. This is accomplished through repetition as computers need to be fed as many images or videos as possible. On the other hand, human vision revolves around light, and, while it and computer vision both have inherent bias, CV can be fooled more easily. Recent computer vision frameworks have been found susceptible to well-crafted input samples called “adversarial examples”. Adversarial perturbations can easily fool DL models in the testing stage. As the susceptibility or liability to adversarial examples becomes one of the major risks, attacks and defenses on adversarial examples are important considerations when developing and applying deep learning in safety-critical environments.

Asif: Modern times are seeing a rise in deepfakes. Can you explain the technology behind deepfakes? How do you see these emerging technologies impacting daily lives?  How can we build more trust around AI?

Sherin: Synthetically generated, highly realistic altered videos, also known as “deepfakes”, continue to capture the attention of computer graphics, CV, and security researchers. Recent advances in these fields and DL have made it increasingly easier to synthesize compelling fake images, audio, and video. The possibilities of the adoption and weaponization of deepfakes are causing alarm in the digital realm.  

I have researched the potential of Generative Adversarial Network-based (GAN) technologies to use in deepfake creation. The GAN training incorporates a generator and discriminator. The generator takes in an input image and the desired attribute to change, then outputs an image containing that feature or attribute. The discriminator will then try to differentiate between images produced by the generator and the authentic training examples. The generator and discriminator are trained in an alternate fashion, each attempting to optimize its performance against the other one. 

Ideally, the generator will converge to a point where the output images are so similar to the ground truth that a human will not be able to distinguish the two images. Thus, GANs can be used to produce “fake” images that are very close to the real input images. GAN techniques such as AttGAN, StarGAN, and STGAN are primarily partial face manipulation methods, whereas PGGAN and StyleGAN2 can be used for full-face synthesis. 

Such sophisticated doctored videos do threaten our political, legal, and media systems. The mere existence of deepfakes undermines confidence and could destroy our trust in society. We will need to develop novel forms of consensus, new ways to regulate the use of deepfakes, and new ways to agree on social situations based on alternative verification forms of trust.

Asif: How are deepfakes currently being detected by AI researchers? How can AI researchers improve their detection methods?

Sherin: Current frameworks mainly focus on soft biometrics, CV, and DL algorithms to detect deepfakes. One of the early research directions made use of detecting eye blinking, a physiological signal that is not well-presented in synthesized fake videos. However, sophisticated forgers can now create realistic blinking effects with post-processing and more advanced models.  

Other research applied to detect deepfakes is based on inconsistent head poses and facial image warping defects. Few other works aim to improve the generalization ability of a CNN (Convolutional Neural Network) forensics model.  It adds an image preprocessing step before training to force the discriminator to learn more intrinsic and generalizable features. These frameworks look at detection of Gaussian blur, shading artifacts arising from illumination estimation, and imprecise geometry estimation of the facial features and missing reflection. CNNs, RNNs and LSTM-based frameworks and pre-trained models such as VGG, inception, Xception, ResNets have also shown promising results to detect deepfakes. 

With the recent release of large scale deepfake datasets with additional annotations, I hope this research will continue to advance and help alleviate the problem. Hopefully, this research will lead to larger high-quality deepfakes datasets in the future. 

As a future area of research, it would be interesting to make use of transfer learning ability to further generalize current models as well as the use multi-modal techniques.  Another potential area to consider is localization of manipulated pixels in GAN-generated fake images. This can be done by either proposing better localization methods which might prove useful to unseen GAN methods through exploiting the imperfection of upsampling methods. With advanced GANs such as AttGAN, StarGAN and StyleGAN2, it would be also helpful to visualize the fake texture in each image and classify them according to different GANs and upsampling methods.

Asif: What is explainable AI (XAI)? Tell us how it works in the context of malware.

Sherin: The cybersecurity industry leverages ML and DL techniques to combat ever evolving cyber threats such as malware. While ML and DL models have become increasingly important for decision-making and are making impossible feats possible, these models are, in essence, a “black-box” as the process that models use to make predictions can be hard for humans to understand. XAI proposes the industry make a shift towards more transparent AI by creating a suite of techniques that produce more explainable models whilst maintaining high performance levels. XAI allows DL models to be more transparent by providing explanations of their decisions and allowing users, customers, and stakeholders to gain insight into the system’s models and decisions.  These explanations are important to ensure algorithmic fairness, transparency, and privacy, as well as to identify unconscious bias, data drift, model decay, and potential problems in the training data. XAI also ensures that the algorithms perform as expected. With XAI, domain specialists, analysts, and customers can understand and analyze actions and predictions, even of the most complex neural network architectures. The challenge with XAI lies in balancing the benefits with exposure of any feature-based confidential and intellectual property; as XAI improves, the demarcation lines may get blurred.

Asif: Can you name some books, courses, or other resources that have influenced your thoughts the most? 

Sherin: The books Computer Vision by Richard Szeliski and Pattern Recognition, Machine Learning by Christopher Bishop are great starting points for learning fundamental concepts. These have helped me gain a comprehensive understanding of the disciplines of CV and ML. The pattern recognition field has undergone substantial development over the years. In addition to recent papers, books such as The Elements of Statistical Learning by Trevor Hastie, Robert Tibshirani and Jerome H. Friedman, and Convex Optimization by Stephen Boyd are also helpful in understanding and supporting research on ML and related fields. If anyone prefers specifically learning about specialized topics in CV such as t3D CV and methods related to inferring geometry from multiple images, the books Multiple View Geometry in Computer Vision by Richard Hartley and Andrew Zisserman and Introductory Techniques for 3-D Computer Vision by Emanuele Trucco and Alessandro Verri will come in handy. 

Asif: What advice would you give to machine learning students who want to jump into the industry?

Sherin: ML and DL form the crux of AI. In addition to these books, I would highly recommend taking a formal course in the area of statistics, ML, and DL. There are more advanced math courses on optimization techniques, which are very good if you are interested in specializing in that area.  One can also pursue online certifications, boot camps, and massive open online courses (e.g., Coursera, edx, Udacity) on the specialized topics.

Once the fundamentals are obtained, start practicing on open-source datasets, IEEE contests, and competitions to gain some practical feedback and experience. Additionally, try to attend local meetups or academic conferences. This will help not only with staying up to date with the latest research, but also provide an opportunity to meet more experienced folks in the area. Most importantly, it’s essential to understand that one needs to constantly keep learning, as learning and innovation go hand in hand.

Asif: What are your views about MarkTechPost.com?

Sherin: MarkTechPost is a great information resource community for both aspiring and experienced data science professionals. It provides the latest research updates in the area of ML, DL, and data science and does have great materials under AI paper summary and University Research articles. With free tutorials on AI and video lectures, I think it will be a helpful resource for many aspiring data scientists as well. I hope the community keeps growing towards building and spreading awareness on next-gen data science ecosystems.

Acronyms used: 

AttGAN, PGGAN and STGAN stand for Attribute GAN, Progressive GAN and Selective Transfer GAN. RNN and LSTM corresponds to recurrent neural network (RNN) and Long short-term memory (LSTM). Both fall under the class of artificial neural networks used in the field of deep learning.

The post Exclusive Talk with Sherin Mathews, Senior Data Scientist at McAfee appeared first on MarkTechPost.