I recently caught up with Atif Kureishy, Global VP of Emerging Practices at Teradata, during the 2019 edition of the NVIDIA GPU Technology Conference, to get a deep dive update for how Teradata is advancing into the fields of AI and deep learning. Based in San Diego, Atif specializes in enabling clients across all major industry verticals, through strategic partnerships to deliver complex analytical solutions built on machine and deep learning. His teams are trusted advisors to the world’s most innovative companies to develop next-generation capabilities for strategic data-driven outcomes in areas of artificial intelligence, deep learning and data science.
insideBIGDATA: Can you give us a brief overview of your role with Teradata and the company’s current direction?
Atif Kureishy: I’ve been at Teradata for two and a half years now. In that role of promoting AI and deep learning, my teams do three things. Number one, we work with Fortune 500 companies to advance their deep learning capabilities. We also have a separate ML team. My team is focused on deep learning, leveraging a lot of open source, building solutions to solve problems and drive what we call “answers or outcomes.” The second thing we do is we take all of that experience, lessons learned, frameworks and toolkits, approaches, and IP, and work with our product teams to say, “Look, here’s what we did. Here’s what was successful. Here’s what’s not. Here’s what’s hype. Here’s what’s in the applied research community. Let’s consider how we would bring all that into the technology portfolio of Teradata and shape the roadmaps and shape where we’re going in the future. As part of that we’re looking to GPU accelerate and provide deep learning capabilities into next year. And item number three is working with media, industry analysts, with the community, conferences to help people understand where the hype is, where the reality is, and that takes a big portion of our time. We meet with customers to help them understand what the state-of-the-art is, where the opportunities are, and how to get there.
So that’s generally me and my team. I have a global remit, so I see all geographies. China, Japan, Europe, North America are probably the areas that we focus on most with Singapore, Australia a little bit, but really, those are the territories that have the most investment and spend of our customers.
The industry that we focus on most, not surprisingly, is financial services, in particular banking. But we also deal with telecommunications, media/entertainment, retail, and manufacturing. Those are always the early adopters of tech, so that’s what we’re seeing.
insideBIGDATA: How is Teradata changing their message in light of changes in the tech industry these days?
Atif Kureishy: The reason I came to Teradata is because the company was in the midst of a strategic transformation and I was enamored with that transformation. We’ve done MPP really well, serving the largest enterprises. Over time those customers get acclimated and they say, “Well, it’s great. We keep the lights on, but I want to do more with this data. You’re helping me with the BI reporting, but increasingly I want to apply more advanced analytics, and I want to get into the predictions.” So Teradata realized it needed to be more of an analytics platform, and enable our customers to do modern data science and modern analytics.
To do this we needed to bring in new talent, and also transform the product. It’s more than just a SQL-based engine. It’s appreciating that you have different types of analytics and computation that you need to apply on that data, relational data. That’s semi-structured, unstructured, image, voice, etc. Increasingly our customers want to apply multimodal types of analysis. Teradata is now on this pivot to approach this transformation.
Recognizing that if you’re moving to get out of the SQL data infrastructure game, then the buyers change and the marketing and the go-to market changes. Who has the dollars for AI? It’s really the line-of-business that has the dollars for AI along with the AI agenda. In order to engage with the senior executive in the business, you try to help that leader with any number of outcomes, like people prediction or anomaly detection or yield optimization. You need to speak their language, understand their business, but ultimately bring data together and apply machine and deep learning to those problems. Narrowly, I will say I am focused on specifically deep learning because we focus on new value creation. If you can apply deep learning in the enterprise and solve the problem of AI explainability then you can do things you’ve never been able to do before. For instance, in working with a very large manufacturer, doing high pressure hose fabrication, their default rates of scrap was off the charts. They had something like 30% of what they manufactured end up as scrap. When we took a closer look at that, we found they had sensors that were misaligned and as a result they had many false positives, and they had a lot of teardown from the quality teams for high-pressure hoses that were completely fine. So if you can take process data, and sensor data, and apply neural networks to increase the accuracy of when you predict a defect, the net effect is hundreds of millions of dollars.
In another instance, you can start to understand customer behaviors when they come into a store by using at computer vision techniques. This is what Amazon is doing with their cashier-less stores. You have sensors in really high definition cameras, and you can start to track how customers traverse through the store, and start to appreciate how long they dwell and queue and different things like that. You can optimize the layout of your store. You can put digital signage in the right places. You can optimize your staff deployment. We’ve been working with the largest retailers for a very long time and many are interested in this technology. How do we change the game and allow them to do this? It’s all predicated on data, but obviously, you need that data, and you need to analyze that data. So that’s why Teradata recommends that they need to get into that sort of analytic space and into the predictive space, hence the transformation.
insideBIGDATA: That’s a pretty big refocus. What was the time frame of this pivot for Teradata?
Atif Kureishy: It’s been over the last three years. We’ve been internally focused on it, but if you’ve seen the sort of rebranding and refresh of our go-to-market, we’re focused on pervasive data intelligence. Let me break down those three words. “Pervasive” in the sense of you need to be able to process all this different types of machine data, log data, structure data, curated data, etc. and process it where it is – in the cloud, on-prem, in object stores, in relation stores. Increasingly if you do analytics on samples of data, you don’t really get the full view. Scale becomes a big issue and Teradata has always been about performance at scale. The second word, “data,” is our legacy. Finally, “intelligence” is the appreciation of artificial intelligence, and the way of prediction and better insights and understanding is on that data, at scale, everywhere.
So in a lot of ways, it’s not a dramatic pivot. We’ve been doing distributed algebra and analytics on Teradata forever – the SQL-based capabilities. Now you’re talking about linear algebra, discrete math, calculus, differential equations. You’re applying more sophisticated types of math. When you talk about deep learning, you’re applying more sophisticated math on that data. But what everyone struggles with is how you do that at scale. We’ve got the scaling part figured out. You need to reach beyond just algebra into geometry, which is what you need – Euclidean geometry in a lot of computer vision problems. But at the end of the day, it’s just math at scale on data, and so that’s what we’re talking about.
insideBIGDATA: And that’s what NVIDIA brings to the table, yes? How are you guys working with NVIDIA?
Atif Kureishy: Absolutely.
We’ve been a partner with NVIDIA for about one year, part of the Services Delivery Program (SDP). If we engage with the customers and help them solve deep learning problems, that’s going to push computation on the GPUs. So obviously, that’s very harmonious.
Coming up next year, we’re actually putting compute into our Vantage platform. You’re running workload on the Teradata Vantage platform, and that data and computation will be processed on GPUs for training, and serving the inference side. Ultimately, you’re solving answers and problems for our customers. Our 2019 focus is Vantage. We have all the computation and data, along with Teradata Everywhere, AWS, and Azure. But let’s forget about all of that. The idea is if you can deliver this in an “as-a-service” manner which really means in a more consumable way to align a business executive.
We can do it in a much more innovative and creative way using machine and deep learning. But we’re not going to bring all that complexity. We’re going to give you a subscription or some straightforward consumption-based method offering dashboards, data pipelines, ML frameworks, data labeling/annotation schemes, and GPU infrastructure. Every enterprise leader in the business wants all of that sophistication without all the complexity, so that’s increasingly what we’re focused on.
insideBIGDATA: What’s the timeframe for these solutions?
Atif Kureishy: It’s an evolution. You’ll see this carrying along a multi-year strategy. A lot of folks are doing this in the cloud, so we embrace those partners where it makes sense. But the Fortune 100, what we call “megadata” customers because of data gravity, privacy, security, etc. You have to allow them to get to the cloud and that’s a part of our Teradata Everywhere strategy. You also have to allow them to do analytics at scale in that same Teradata Everywhere environment. By the way, deep learning is just an evolution of ML. ML is just an evolution of some of the other modeling and simulation techniques that we’ve been using. So you have to take customers on that path.
It’s available, or will be available, on AWS and Azure, and on managed cloud. So those things are available now, so folks can come on board now, and then when the deep learning capabilities come out, they’ll have access to that technology as well. It’ll be part of a first class environment with Vantage. The idea is that we’re going to take them on that journey, and be there for them when they need it.
insideBIGDATA: Can you describe a particularly use case?
Atif Kureishy: Yes, there were some creative applications of deep learning at Danske Bank with a variety of transactions involving issuing bank, and receiving bank. We decided to extend and add new features around everything else we know about the transactions, such as IP addresses, Mac addresses, and other derivative information. Then we observed that these transactions occur over time, so we were actually looking at sequences of transactions rather than individual transactions. A lot of machine learning approaches today look at a transaction in isolation in order to do comparative analysis and anomaly detection. But we were actually looking at sequences of transactions so there’s better signal in that detection.
So we took the sequences arranged over time and we turn that into a model to emulate pixels on an image. We literally took those transactional features and then did some spatial correlations model techniques and we turned it into image.
Then we applied convolutional neural networks (CCNs) to the image and that became a best-performing method. We did time-aware LSTMs and other types of recurrent neural networks (RNNs). The derivative benefit of this approach was that the auditors and regulators could actually see fraud visually. We showed this kind of pixelization where the intensity of a pixel would actually demonstrate fraud. They got it, and then applied some other techniques to recognize attributes that contribute to the classifier of false deny or approve. This was enough for us to understand what these black box models are doing.
In the end, this solution was an ensemble of six different techniques. We had some logistic regression approaches, some boosted trees, and some other GLMs. Then we used a deep neural network. It was such a dramatic improvement. We worked with them to build their data science capabilities so that they could support this in the future, and that’s why it was such a transformational effort.
insideBIGDATA: Well, this has been great. I appreciate the opportunity to get a Teradata update.
Atif Kureishy: My pleasure.
Sign up for the free insideBIGDATA newsletter.
Speak Your Mind