Memory-Prediction Asymmetry

I think there are unresolved philosophical issues around probabilistic interpretations where it's not clear a priori what the space of possibilities is. Even if we could pin it down in an instant, that space is constantly reshaped in the brain as it adapts to its environment.

"brown dog" surely has some probabilistic content but it's hard to locate where it is because of the enormous complexity of the space of "possible situations in which a brain can perceive a brown dog".

The language (Recognizing vs Generating) sounds awfully close to deep learning, and both are very hot subjects in the last 5 years. A GAN is basically made of a generator coupled to a recogniser. In reinforcement learning there have been promising results by coupling an agent with an internal world model capable of rolling out possible outcomes.