Over the past 18 months -Since my son's birth until now- I keep asking to myself: How does learning work inside that tiny yet everyday more clever brain? Beyond all the chemical mechanics and immense complexity, can one, out of mere observation deduce "simple" algorithms that describe a good amount of what it takes to learn/think as humnas do?
I would like to entertain the following line of thinking: What if our thoughts can be represented as the result of a chain of predictions? The deeper in the chain the more complicated the thoughts can be.
An experiment
Design an algorithm that from a given video, takes a frame and tries to predict/anticipate what the next frame will be like. Also, design that algorithm in a way such that it can continiously and unsupervised; learn from the accuracy of its predictions. Lets call this algorithm the Anticipation Block.
Lets asume that we manage to design such Anticipation Block as desired. What happens if we stack/chain Anticipation Blocks with the objective of predicting further and further into unseen frames? What kind of features will these chains learn after hundreds or thowsands of hours of video running through them?