François CholletVerified account

It has simply not yet shown to be necessary, or even useful. It actually seemed like a more attractive avenue when we knew less and our models performed worse.

ML models don't attempt to emulate human cognition, and they're solving a different problem than embodied cognition in the first place, with different constraints and different degrees of freedom.

If your input is a static image that you're trying to classify, that's a very different setup than being an embodied agent immersed in a dynamic world subject to cause and effect. In the former case, processing all the information available in one go is actually more effective

You can take context into account without active perception. Active perception only becomes really useful in a dynamic world where it's possible to formulate & test hypotheses (requires a time component)

The current deep learning standard for implementing context-awareness is "neural attention" (cf Transformers), perhaps you know about it. It has very little in common with active perception though.

The fact is that hardly any ML model takes "the world" as an input (complete with time, cause & effect). Only static snapshots of it. Ultimately this is why active perception hasn't taken off. If all of AI was cognitive developmental robotics it would be a different story.

Thanks. This is fascinating. I need to brush up on neural attention from an AI perspective, esp since I study human attention & am curious about the overlap. Curious about what challenges lie ahead for ML models & whether they will entail a need to better emulate human cognition.