AAAI 2020: Unsupervised Deep Learning and AI that can reason

Source: Artificial Intelligence on Medium

Yoshua Bengio: Consciousness prior and a world model

Yoshua talked about his efforts in achieving human-level AI. In his words, “There is no completely general intelligence — these are always some inductive bais and priors”. Simpler fewer priors are more likely to be discovered and can be combined to understand a broader set of contexts.

He talked about the possibility of developing a model of the world (or world event) and make a neural network learn it. This can help deep learning learn system 2 in terms of system 1. The concept of System 1 and System 2 came from Daniel Kahneman. In his book Thinking, Fast and Slow, Daniel Kahneman describes System-1 as fast, automatic, frequent, emotional, stereotypic, unconscious, anchored in sensory perception in humans and System-2 as slow, effortful, infrequent, logical, calculating, conscious. rational, sequential, and expressible with language.

System 1 operates automatically and quickly, with little or no effort and no sense of voluntary control while System 2 allocates attention to the effortful mental activities that demand it, including complex computations. The operations of System 2 are often associated with the subjective experience of agency, choice, and concentration.

Deep learning at some level is able to do automatic perception like system 1 and the challenge is how to extend DL from system 1 to system 2. If we need some sort of consciousness prior to representing high-level concepts and use DL to predict the value of the various priors. If these priors can be combined in some way to do some real human level reasoning and understanding?