Gang Fang's Blog

Bengio’s world model

Bengio’s concept of world model is a generative model that learns the abstract structure, or abstraction, of x. And this is similar to how human learns and understands the world. This is powerful because this enables training with far less data (because we don’t operate at the low level anymore) and better out of distribution generalization.

Rather than simply focusing on the variables like x, it might be more important to focus on how they relate to each other, which represent the causal mechanisms humans use.

And this causal mechanisms are represented in graphs.

Also, the world model then can be represented by joint distribution of high-level variables. In turn, these joint distributions can be modeled by sparse factor graph.

Why do we need the factor graph to be sparse?

Because we only want to capture relationships among HIGH LEVEL variables. The sparse factor graph won’t be able to predict the next pixels given previous 4, but it can predict the next word given previous 4 in a language (high level)

Click to access NeurIPS-11dec2019.pdf

Bengio has been thinking about this for 20 years.

QUESTION: what is the place of a world model in the GFlowNet’s landscape?

The reward function R used to train a GFlowNet resides in a world model.

Separating world model from the inference machine, an anecdote

My shoulders were in pain today so I listened to, instead of watching, a YT video on Picasso. A thought popped up that in human history passing of information has been done primarily through auditory means but not visual. I held this thought and then moved to lay down on a yoga mat because I didn’t need to look at the screen. Then, I picked up this thought and further analyzed it: visual forms of information passing could trace back to the beginning of cinema, which processed TVs and video streaming while auditory storytelling has been existing for thousands of years. Then I realized, the making of static pictures as a mean of storytelling came before cinema and traces back to prehistoric time, similar to the verbal counterpart.

After this exercise, I became aware of the fact that I was unintentionally verifying my thought/assumption against the world model that I have, and facts and information were retrieved from the world model to answer the question being held in my conscious mind to complete the inference, which eventually led to a judgement of the thought in focus.

And this subjective experience mirrors the new architecture that Bengio has been proposing.