Interview with Professor Hector Levesque, University of Toronto

Excerpt

Hector Levesque is Professor Emeritus in the Department of Computer Science at the University of Toronto. He worked in the area of knowledge representation and reasoning in artificial intelligence. He is the co-author of a textbook and co-founder of a conference in this area. He received the Computers and Thought Award in 1985 at the start of his career, and the Research Excellence Award in 2013 at the end of his career, both from IJCAI (the International Joint Conferences on Artificial Intelligence).

What makes commonsense reasoning so special? It is neither logical nor illogical but demonstrates systematic patterns, how would you characterize it?

What makes commonsense reasoning so special is its range and power, coupled with the limitations it shows. I would characterize it as a form of computation over symbolic representations of knowledge.

How far are we from a general theory of commonsense reasoning?

At a coarse level, we can easily see a lot of the components of commonsense reasoning in logic and probability. But at a detailed level, where we would want to predict exactly where commonsense reasoning succeeds and fails, I think we have a long way to go.

From your perspective - what are current limitations of AI approaches to explain commonsense reasoning?

To me, AI tends to overemphasize the knowledge that needs to be be represented and underemphasizes the procedures that will need to manipulate these representations. We tend to focus on simplistic ideas of what these procedures will need to compute.

Nowadays machine learning with neural networks seems to be the dominant technology in artificial intelligence. Do you think it helps to understand or mimic human/commonsense reasoning? How far?

Deep learning does a terrific job in duplicating and even surpassing certain human cognitive skills. But in my opinion, commonsense reasoning is not one of them.

Hilbert has proposed twenty-three problems that were responsible for large progress in mathematics. Do you see similar challenging problems for cognitive reasoning?

It’s not clear. There have been ”challenge” papers at AAAI, but the startup cost in tackling them might be too high (unlike in mathematics). I proposed the Winograd Schema Challenge, but I’m not at all confident that people will be able to use it to make real progress.

For some problems like the Wason Selection Task there exist more than 16 theories how humans reason - which is devastating for any science. One answer to eliminate theories is to come up with criterias and benchmarks. What do you expect of good cognitive theories?

The Wason Selection Task is an excellent puzzle to try to explain, but our current cognitive theories that would account for its effect seem to be too weak to be definitive. My feeling is that there is not enough attention to the weaknesses and limitations of commonsense reasoning.

Do you think formal logic is an adequate method for modelling cognitive reasoning?

It is a simple normative theory that explains a lot, but is too simplistic to account for the limitations of human reasoning.

What is the status of existing systems for commonsense reasoning?

Not enough space in the margin here...

Is there a chance to enhance neural network machine learning systems by explaining components?

Deep learning systems are not well-equipped to have explanation components tacked on after the fact. Even in very simple cases, the approach is not well-suited to align with a human-like understanding of the world.

Is there hope to learn form neuroscientits about how to handle large amount of knowledge and how to use it for reasoning?

I think the main ideas will come from elsewhere. But neuroscience provides useful constraints on what accounts of reasoning can be made to work in the brain, somewhat like electrical constraints on what can be computed.

What would be a question, you would like to answer?

Here is a question I would like to have answered (by somebody who is not retired): How do people understand free-flowing English text so well in real time, including resolving ambiguities, pronoun reference etc? …