Intel/Saffron AI Plan Sidesteps Deep Learning

　　Intel’s $1 billion investment in the AI ecosystem is one of the well-publicized talking points at the processor company. The Intel empire boasts a breadth of AI technologies it has amassed by acquisition and Intel Capital investments in AI startups.

　　The acquired companies seemingly useful to Intel’s AI ambitions thus far include Altera (2015), Saffron (2015), Nervana (2016), Movidius (2016) and Mobileye (2017). Intel Capital has also fattened its AI portfolio with startups Mighty AI, Data Robot, Lumiata, CognitiveScale, Aeye Inc., Element AI and others.

　　Unclear is how Intel is going to stitch all this together.

　　With AI innovation still in its early days, Intel’s apparent scattershot approach to AI strategy might be justified. We might still have to wait a while for a more coherent narrative to emerge.

　　Intel has talked up its AI hardware portfolio more often than its overall AI strategy. Typically, Intel announced this week that it will shipbefore the end of the year the Nervana Neural Network Processor (NNP), formerly known as “Lake Crest.” Naveen Rao, formerly CEO and cofounder of Nervana, and now Intel’s vice president and general manager of AI products, describes NNP as featuring “a purpose built architecture for deep learning.”

　　Intel has other ammunition when it comes to AI chips, including Xeon family, FPGAs (from Altera), Mobileye (for automotive) and Movidius (for machine learning at the edge).

　　However, Intel has been reticent about AI applications or exactly which fields of AI where it will focus. AI is a realm both broad and deep. Among Intel’s sprawl of acquisitions, the biggest mystery might be Saffron.

　　It was hard not to notice earlier this month when Intel announced a new product called the Intel Saffron Anti-Money Laundering (AML) Advisor. The product isn’t hardware, although AML apparently runs on Xeon processors, but a tool for investigators and analysts to ferret out financial crimes.

　　Earlier this week, EE Times had a chat with Elizabeth Shriver-Procell, director, financial industry solutions at Saffron Technology, to learn about the AI technologies behind Saffron’s product, and what she sees as Saffron’s gain in becoming an Intel company.

　　Mainly, though, we wanted to know what a long-time financial crime-fighter like Shriver-Procell is doing inside the world’s largest CPU company.

　　EE Times: Tell us a little bit about yourself. I hear you are an expert on financial analytics, working at various companies and agencies including the Treasury Department.

　　Shriver-Procell: I am a lawyer, with the focus of my work on fighting financial crimes. I’ve worked at international consultancies and various financial institutions. Most recently, I was at Bank of America. I joined Saffron earlier this year. Yes, I also worked at the U.S. Treasury as a program manager for analytics development.

　　EE Times: So before coming to Saffron, did you use Saffron’s products?

　　Shriver-Procell: Some organizations I was associated with — including some clients at consulting companies — have used Saffron. I’ve been intrigued by the platform, so when this opportunity came up, I took it.

　　EE Times: So, what exactly does Saffron offer?

　　Shriver-Procell: Saffron was always sold and marketed as an ‘analytic platform’ customizable for broader applications. Users include supply chains, banks and insurance companies.

　　EE Times: With the launch of Intel Saffron Anti-Money Laundering Advisor, has anything changed in Saffron’s platform approach?

　　Shriver-Procell: We’re now rolling out specific products for specific applications.

　　Different branch of AI

　　EE Times: I suspect the primary reason for Intel to acquire Saffron was more to do with getting their hands on Saffron’s AI technologies, rather than solving financial crimes (although it’s a worthy cause). Tell us a little bit about what kind of AI expertise Saffron has designed and uses for what you do. And how is that different from other AI technologies?

　　Shriver-Procell: At Saffron, the AI technology we use is called Associative Memory AI, which is a different branch of artificial intelligence from Deep Learning. Associative memory AI is very good at looking at a large volume of data – and a high variety of data – and discern signatures or patterns out of databases that are so far apart. It unifies structured and unstructured data from enterprise systems, email, web and other data sources.

　　EE Times: Give us examples.

　　Shriver-Procell: Take the example of a banking customer named Mary. Mary goes to London every other week and shops at Liberty store. John, who lives in a different country goes to London about the same time when Mary is there, and does something entirely different. Is there any relationship between the two? What are the commonalities between the two? Can we take a look at their IP addresses? Do we find any similarities in their log-in patterns? Is there anything that shows if any nefarious activities are going on there?

　　EE Times: So, the point is that Associative Memory AI can take a look at so many seemingly unrelated databases at the same time.

　　Shriver-Procell: Not only that, it gets the job — otherwise very time-consuming — done very quickly. While it takes a lot of training for Deep Learning to work, Associative Memory AI does not need to be trained. This AI does rapid, one-shot learning. It’s a model-free AI.

　　EE Times: In the press release, you talk about Saffron’s “white box AI.” Please explain.

　　Shriver-Procell: By “white box AI,” we’re talking about transparency. We can explain how we have arrived at a certain conclusion. In the past, financial institutions acquired a model-based, vendor-supplied solution for fraud detections. We call it a “black box” because users have no idea how their software has worked inside the black box. When regulators ask financial institutions how they came to a conclusion, they can’t really explain it. They can’t see what’s inside the black box, and they can’t tell if it was working properly.

　　In highly regulated industries, it’s critical for financial institutions to be able to provide transparency in their data.

　　EE Times: Interesting. That sounds like almost the opposite of Deep Learning AI. Some safety experts worry that when Deep Learning AI deployed in autonomous vehicles makes a certain decision turning a corner, for example, carmakers can’t explain why the AI made its decision. The lack of transparency in the learning process makes it tough for carmakers to validate the safety of autonomous cars.

　　Shriver-Procell: I think it’s important to recognize that there are different approaches to AI. When Intel’s CEO talks about unlocking the promise of AI, he says that we may try new things. We need to explore new learning paradigms.

　　EE Times: Do you see that any of those different branches of AI converging at one point?

　　Shriver-Procell: I think they are complementary. As we see a growing trend for blending of applications, I think multiple types of AI will be able to address the needs presented by a spectrum of applications.

　　EE Times: Tell us more about your new products.

　　Shriver-Procell: As I said before, Saffron always sold its product as a platform. Now, as we are finding specific needs in specific market segments, we’ve decided to roll out a specific solution as a product that can meet the market’s challenges.

　　Saffron has always held a very strong position in the financial market backed by its experience in finding financial crimes. By unifying structured and unstructured data linked into a 360-degree view, we can make sense of the patterns found across boundaries wherever the data is stored.

　　We also announced that the Bank of New Zealand has just joined the Intel Saffron Early Adopter Program. This is designed for those institutions interested in innovation in financial services by taking advantage of the latest advancements in associative memory artificial intelligence.

　　EE Times: What do you think Saffron has gained by becoming an Intel company?

　　Shriver-Procell: The benefits of joining Intel are great. We’re talking about serious problems that large financial institutions are fighting. In order to be able to support them, you need all the power and support that a large corporation like Intel brings to bear. We also need the full support of Intel as a technology partner as we create new capabilities and applications on the Saffron platform, and make them scalable and extensible. As AI rapidly advances, you can’t overlook the significance of exploring new things, new ways to do AI.

　　AI in its infancy

　　After the interview with Saffron, EE Times got in touch with a few analysts to see how they view the state of AI technology development.

　　Jim McGregor, founder and principal analyst at Tirias Research, observed, “There are many different types of learning (supervised, unsupervised), different types of digital neural networks (deep learning; holographic associated memory — also referred to as just associative memory, inference models), different hardware solutions for AI (CPUs, GPUS, DSPs, FPGAs, TPUs, Quantum processors), and a plethora of different software frameworks. So, mapping out all the AI solutions is like mapping out a tree that has new branches springing out every day.”

　　Paul Teich, principal analyst at Tirias Research, concurred. “New classes of learning and AI algorithms are still emerging at a frightening rate.” He added, “That means we are still fairly far away from locking in efficient full-custom silicon. General purpose silicon rules during times of radical change. That is why GPUs, FPGAs, and coprocessor style matrix math accelerators (NVIDIA's Tensor Core and Google's TPU2 are in this bucket) will dominate until we get farther down the road in selecting best in class algorithms and best practices for model development and deployment.”

　　Do we see some of those different branches of AI in the future working together?

　　McGregor said, “This is a great question.” As he sees it now, “Most of the effort is being put on Centralized Intelligence and Hybrid Intelligence, where everything is done in the cloud for split between the cloud for learning and the edge devices for inference. A few companies like Microsoft are working on distributed intelligence where the intelligence can be spread amongst multiple resources, such as data centers for deep learning.”

　　In his opinion, “The future will require Collective Intelligence where all these intelligent solutions work together. We do see the future as being one of collective intelligence.” But he noted, “When and how we get there has yet to be determined. Which solution has priority? What do you do when these solutions do not agree? How do you collectively share information between drastically different frameworks and neural networks (even creating two neural networks that look the same using the same data is next to impossible)? These are all issues that will have to be worked out.”

　　McGregor added, “I'm not surprised to see Intel starting with [Saffron], because the financial industry will be one of the industries that drive us toward collective intelligence because of its importance to the global economy.”

　　Now that Intel is offering Saffron’s Anti-Money Laundering Advisor as “a product” in the financial market, does this mean that Intel is taking a step — somewhat akin to IBM — toward a “service business model” rather than just sticking to the chip business?

　　McGregor believes it is. “Intel has done this before and tends to swing back and forth between being a solutions vendor and a technology vendor, but in the case of AI, you almost have to be a solutions vendor because of the need for both hardware and software, and Intel has invested in both.”