IBM Corp. is justly famous for pioneering work in artificial intelligence, even decades before its Watson computer beat a couple of “Jeopardy” champions. But in recent years, its work has been eclipsed at least in the public imagination by new AI-driven speech and image recognition services and self-driving cars from companies such as Google LLC, Microsoft Corp. and Facebook Inc.

Today, the veteran tech company is highlighting in a blog post a few of the more than 100 papers its researchers published this year on AI and machine learning. They’re broadly focused on advancing the state of the art, scaling up AI systems and trying to increase trust in AI systems that have come under fire from governments and consumers alike for their lack of transparency and problems of bias.

SiliconANGLE spoke with Jeff Welser, a vice president and lab director at IBM Research Almaden Research Center in San Jose about a few select examples of what IBM’s working on, challenges it’s confronting in AI work and what its researchers think is coming in 2019 and beyond. This is an edited version of the conversation.

Q: What are the prospects for deep learning or other AI systems that can learn like humans do, from relatively few examples?

A: This is a challenge. But as we think about the way people learn, as a kid you’re given a few examples of a car and your mom says that’s a car, providing a couple of labeled examples. But you actually do see many, many, many cars, so you do get lots and lots of unlabeled examples, and that helps you define all the parameters and features that make up a car.

But at some point, you do see something that’s out of the ordinary, like really large, and you’ll ask “Is that a car?” and mom says, “No, that’s a truck.” So instantly, when you get to that example, you see when the parameters get to a certain level, I now know that’s a truck and you immediately see lots and lots of examples of trucks and your brain is really good at categorizing the two.

That’s really what we’re going after, one-shot or few-shot learning. We’re reducing the labeled data and helping the system extrapolate so it can recognize new things that are related to it. Ultimately, you want the system to know to ask. The way we handle it now is we try to do things like provide confidence levels. But that’s different from it knowing what it doesn’t know.

Q: Are there other ways of reducing the need to fuss with so much data?

A: We’re also looking at ways we might be able to autolabel samples of lots and lots of images. Or even systems that can help humans do the labeling more quickly by clustering images that might be alike and asking, “Are these right?” We do that a lot in natural-language processing.

Q: What methods are you looking at beyond deep learning?

A: There’s a renewed interest in rules-based analysis of unstructured data. One of the challenges of neural nets is it can be a black box. We don’t always understand exactly which features has the neural net learned to allow it to understand what a car is, for example, which makes it difficult to know when or why it makes a mistake. One thing is can we analyze neural nets better to understand what features they are actually making decisions off of.

But the second thing is, should we go back to some rules-based work? Try and teach the system that there are common-sense things, like water is wet, water will stay in a bowl.

There is some interest in figuring out if you can combine them. Even some of the systems we use do this — System C, a text analytics system, is very much rules-based. It’s augmented by deep learning. One of the things we found was that in some cases, although rules-based systems are harder and more complex to build and maintain, they can be more robust.

Photo: Jeff Welser/Twitter

Q: How else are researchers looking to break out of the black box of AI?

A: One approach is can you look inside neural nets and analyze what neurons seem to be firing on various features that are being presented. Is this part of the neural net responding to corners or edges or variations in movement if it’s video? And then infer from that what is really triggering responses. If it’s triggering something strange, maybe that’s a signal that there’s something we should look into.

I have a group at Almaden working on more speculative work to analyze neural nets and extract the rules that it’s actually using. It’s more challenging but it’s interesting because it might bridge the two.

Q: What are you doing to reduce bias in AI?

We’re doing a lot of testing of neural nets with broad sets of data. It becomes most clearly an issue when you talk about bias. If you have really large data sets, you might not even realize that the data are slightly biased on gender or whatever you’re analyzing

So now we have systems to analyze a data set and see if it’s overrepresented in this particular set of characteristics. It might be that you’ve overtrained on those characteristics. Some of the stuff is easy: If you’re trying to analyze people, make sure you have an equal number of men and women.

It’s not going to be a one-shot deal. Data changes and things it will be exposed to change. Earlier this year, we released an AI bias toolkit — techniques and processes we recommend.

Q: How do you view the role of government in AI? Even Microsoft recently called for the federal government to help determine standards.

A: As an industry we are working together on these standards. We want to make sure we don’t take AI in directions we don’t want it to go. The Future of Life Institute is one group that’s looking at how AI may affect the future of society.

Q: Can we keep humans in the loop in AI?

A: There’s two parts to that. There’s keeping humans in the loop when it comes to training the system. We’re using various tools that help balance what the machine can do at the very large scale we’re talking about and what the human can do to make sure it’s driving in the right direction.

It’s similar when we talk about using the system. The goal is to figure out how AI can automate things in a workflow at very large scale that would be very hard for a person to do. But then it should present the information in a way that the person can do what they’re very good at, which is, first of all, understanding what the goal is, weighing common-sense decisions and understanding if the machine is confident or not confident, whether the person should dig in and figure out how the answer was developed. We want that partnership [with the machine] to be easy.

Q: What trends in AI are you tracking next year?

A: One is that trusted AI will take center stage. We’re already seeing that. It’s going to become less an afterthought. You start from the very beginning with the data you choose to train with. Making sure you’re building the system from the bottom up … so the person getting the answer knows where it came from. Trust has got to be part and parcel of training, deploying and using AI in 2019.

The next one is a little more speculative, one that we won’t solve in 2019 but will make progress on: Causality will take over correlation. Big-data analytics basically identifies correlations, such as [image] features associated with a cat. It’s more important when you’re trying to figure out a disease in a diagnosis, for example. A rooster’s crow didn’t cause the sun to rise, it’s correlated with it, but a switch does turn on a light. But AI today is really bad at determining causation. We want to work on understanding causation in AI.

Q: What projects in this area is IBM working on?

A: We’re working on it in the chemistry area to understand what new molecules might be interesting to go produce — understanding the chemistry so the AI doesn’t suggest just random molecules but ones that might actually work. Similar things are going on in other domains in other labs.

Q: Any other trends for the coming year?

A: One area is the furthest out but where we’ll definitely see progress in 2019 is the whole area of quantum computing as applied to AI. Quantum computing in general has been accelerating and I think we’ll see a shift in coming years away from just cryptography and encryption and more for other applications where it’s likely to have impact much sooner.

Those areas are simulations such as in chemistry, optimization problems such as you might have in the financial industry and increasingly machine learning. In AI, you oftentimes need to build a very high dimensional state, looking at a large set of data on a lot of parameters and then figure out how to separate out the right parameters. Quantum computers are very good at generating high dimensional states. And those problems are highly amenable to the kinds of quantum computers we have now.

Q: We’ve heard more in recent years on the application of AI to consumer problems by companies such as Google, Facebook and Microsoft, but we’ve heard less about IBM’s less consumer-oriented work in AI. Why is that?

A: The consumer side is much more visible and easier to talk about and even implement … like recommending products, it’s a very specific task to do. We’re trying to figure out ways to really advance AI much more broadly. We want to help everybody… whether it’s healthcare, the financial industry or even societal things: better interpreting data from IoT sensors, to better control your factory or your house, or figuring out problems with pollution.

Image: cblue98/Flickr