Can lack of causality halt Artificial Intelligence?

mediumThis post was originally published by at Medium [AI]

Photo by Franck V. on Unsplash

Artificial Intelligence, the science of empowering machines with human-like intelligence, has sparked an inevitable debate of Artificial Intelligence vs Human Intelligence and wild predictions about the future. Thanks to rapid advancements in AI over the last decade, computers have become extremely good at diagnosing diseases, recommending movies and songs, and suggesting replies to emails. To add to this, they are even outplaying humans at complicated strategy games like Go, generating life-like images of imaginary people, and even understanding emotions based on the tone of your speech!

Yet despite having achieved all these majestic feats, AI still has glaring weaknesses.

What does the present AI lack?

Artificially Intelligent systems can be confused with ease by putting them in a situation which they have not experienced before. A self-driving car gets baffled by a situation that even a novice human driver could handle easily. A Deep Learning system designed and trained to perform one task (say identifying cats) has to be laboriously trained all over again to do something else (identifying dogs). In the process of re-training, some expertise displayed by the system in the original task is prone to lose. Scientists working with AI have termed this problem ‘Catastrophic Forgetting’.

Photo by Ehimetalor Akhere Unuabona on Unsplash

All these situations exhibiting catastrophic forgetting have one thing in common: they exist because AI systems don’t understand causation. I repeat — AI systems do not understand causation. Sure they see the association between some and the other events but they just can’t seem to ascertain which things directly make other things happen. It’s as if you know that exposure to the Sun makes sunburns likelier, but you didn’t know that exposure to the sun causes sunburns. Understanding cause and effect is a big aspect of what we call as sense, and this is an area where the AI systems of today seem to be absolutely ‘clueless’.

An AI machine’s ability to spot correlation — e.g., that exposure to the Sun makes sunburns more likely — is merely the plainest level of causal reasoning. This level of causation was good enough to have propelled the AI technique of Deep Learning over the past decade to where it is now. Pretty accurate predictions can be made using this method, provided there is a great amount of data available about the situation in question. A computer can calculate the probability that a patient with certain symptoms has a certain disease because it has learned just how often thousands or even millions of other people with the same symptoms had that disease.

Why is understanding Causality important?

An understanding of causality will significantly up the AI game.

But if the latest research and reports about AI are something to go by, the progress in AI will encounter hiccups if computers don’t get better at understanding causality. If our machines could get around the idea of certain things eliciting others, they would not have to learn everything again, every single time — they could instead take what they have learned in one area and apply it to another. Consequently, if our machines could then go on to wield common sense, they will be highly unlikely to make stupid slips and mistakes and in turn, would enjoy more human trust put in them to make their own decisions. An even more sophisticated understanding of causality will be the ability to reason about why things happened and ask the “what if” questions. A pandemic is spreading thick and fast through a population, what measure would be the most effective to curb it? A driver died while test-driving an autonomous car, was it the fault of the computer system or something else?

In the present scenario, if you wanted a neural network to detect when dogs are jumping, you’d show it many many images of jumping dogs. If you wanted it to identify dogs running, you’d show it many, many images of running dogs. The systems would learn to distinguish jumpers from runners by identifying features that are different in both sets of images such as the position of the head and the arms and much more.

However, experts in the field of AI believe that elementary knowledge about the world can be understood by analyzing similar or “unvarying” features across data sets. Maybe a neural network could identify that leg movement is what causes both running and jumping. Maybe after seeing these and many other examples that show dogs a few feet off the ground, an AI system would eventually realize something about gravity and how it limits the physical movement in dogs. After a considerable period of time and with enough meta-variables that are consistent across data sets, a machine could gain knowledge of cause and effect that can be used across domains. AI will be truly intelligent only after it would have achieved a rich understanding of causality, for causal reasoning may lead to introspection, and that is the very core of human cognition.

“What if” questions “are the building blocks of science, of moral attitudes, of free will, and of consciousness.

Judea Pearl, a Turing Award-winning scholar

Spread the word

This post was originally published by at Medium [AI]

Related posts