Inside AI And Deep Learning

Semiconductor Engineering sat down to talk with Dave Schubmehl, research director for content analytics, discovery and cognitive systems at International Data Corp. (IDC), a market research firm. Schubmehl’s research covers information access, artificial intelligence, cognitive computing, deep learning, machine learning and other topics. He also addressed neuromorphic technology. What follows are excerpts of that conversation.

SE: How do you categorize this market?

Schubmehl: We are looking at it as a broad market. We are looking at something we essentially call artificial intelligence platforms or cognitive systems.

SE: What does that include?

Schubmehl: That includes a lot of different technologies. It includes dealing with unstructured information. This could be text analytics for example. Content aggregation technology is another one, which is pulling in a bunch of different things and trying organizing it. Then, there are semantic networks and link data.

SE: I believe machine learning is a sub-set of this category. What is machine learning?

Schubmehl: Machine learning is essentially a set of algorithms. It’s algorithms that self-modify or change over time, depending on the data they receive. What that modification does is that it simulates learning, if you will. So you feed it the right kind of data. The system gets trained on this kind of data. Then, it learns which attributes are important and not important. It learns to recognize those attributes and act on those attributes.

SE: Can you elaborate?

Schubmehl: Within machine learning, there are different types. There is supervised machine learning and unsupervised machine learning. Neural networks fall into unsupervised machine learning. Essentially what happens is the neural network algorithms crunch on the data long enough to identify patterns and identify a set of attributes associated with those patterns. Over time, it learns which of those attributes are important.

SE: That’s not new, right?

Schubmehl: That kind of approach has been around 30 years, maybe 40 years. What’s changed is that the compute power, and the amount of memory and data, has increased so dramatically that we can actually get good results by running these kinds of algorithms.

SE: What is deep learning?

Schubmehl: Now, people like Google and Facebook call it deep learning. It’s essentially the same kind of neural network algorithms, if you will.

SE: You mention cognitive system platforms. What is that?

Schubmehl: They utilize these different technologies that I talked about to essentially help people build artificially intelligence applications or applications that can make recommendations or predictions. They can identify certain kinds of discoveries or patterns. Those are what we consider to be cognitively-enabled applications or artificial intelligent applications.

SE: Who is in this market?

Schubmehl: We’re calling this the cognitive software platforms market. So, it’s the tools to build these platforms. So it includes things like IBM’s Watson. There is a company that Intel bought called Saffron, which has a cognitive system platform. Tata Computer Systems has their Ignio technology. There are also a bunch of startups, such as Loop AI.

SE: How big is that market?

Schubmehl: We estimate it to be about $1 billion for 2015. We’re forecasting that will grow to tens of billions of dollars by 2020. More importantly, what that is going to do is fuel the development of what we call cognitively-enabled applications. We haven’t officially sized that market yet. But it is expected to be larger than $40 billion by 2020 when you include services and data.

SE: Which companies are involved in machine learning?

Schubmehl: Some of the packages, like SAS, SPSS, Cognos and others, have had some level of machine learning algorithms for years. Now, you are seeing companies like Skytree, Loop AI and Mindmeld. They are taking machine learning to the next level. In addition, Google released its machine learning libraries to open source a few weeks ago. There are probably a half-dozen machine learning libraries out there that people are using.

SE: What are some of the key apps and where is this all going?

Schubmehl: They include computer vision, image recognition and pattern recognition. These have improved dramatically over the last five years. In addition, we will have good visual recognition systems in cars, robots, and even your refrigerators. There will be a lot of advances in compute-based image, voice, speech and sound recognition. The next thing is once you recognize that information, what can you do with it? Then, you start getting into machine learning and artificial intelligent programs. They can take those inputs and actually predict what’s going to happen or make a recommendation. The same thing will happen on your desktops. You will have digital agents that help you with your research.

SE: What happen to AI?

Schubmehl: If you talk about broad AI, or this general-purpose AI, that are popularized by movies like Her or Tony Stark’s Jarvis (in Iron Man), I don’t think we are at the point where we are at broad AI. But if you talk about narrow AI, where there’s artificial intelligence to do a very specific job, we are there already. There is a company, for example, called InsideSales. They are using narrow AI to help telemarketers identify what they need to do to close an opportunity quicker.

SE: I recently heard a presentation from Numenta, which has developed a software technology based on principles of the neocortex. Thoughts?

Schubmehl: Numenta is one of a half-dozen companies that are out doing this kind of work. It’s really the Wild West in terms of the AI market.

SE: Most, if not all, of these types of apps run on general-purpose hardware. Are they doing the job?

Schubmehl: A lot of these applications are able to run effectively on GPUs.

SE: Some say the traditional hardware, such as processors, GPUs and memories, are running out of steam.

Schubmehl: It’s always good to look for new ways of doing things. But you have to remember that von Neumann architectures have well over 50 years of research and 50 years of software layered on top of that. That’s pretty important. If you are going to do something that’s brand new, you may have to re-invent all of that.

SE: So can today’s hardware and chips keep up?

Schubmehl: There are going to be advances. Companies like Nvidia are doing it. They are able to parallelize GPU processing.

SE: The neuromorphic chip and systems community is going after many of the same apps being done on traditional hardware. How do you see neuromorphic technology shaping up?

Schubmehl: The goal is to try to develop a computing system that works more like a conventional biological brain. They are fundamentally getting away from von Neumann-type architectures. You have multi-input systems and multi-output systems. The reality is that you may have to invent completely new operating systems, compliers and languages for all of these things to take advantage of them.

SE: So does neuromorphic technology have a future?

Schubmehl: It’s still in the very early days of these neuromorphic systems. It’s tantalizing. The question is what is it going to take to commercialize it.