Robin HansonVerified account

Look into convolutional neural nets. They are "inspired" by the layered organization of the visual cortex.

I'm aware of that. But a design inspiration is not the same thing as a set of design lessons.

Investigation of the architecture of visual cortex has played a large role in the development of deep learning and convolutional neural networks. So far, most AI research therefore mostly has analogies to the functionality of local cortical circuits.

Many people claim their designs are inspired in part by brains. That's different from general lessons we can draw.

The visual cortex inspiration case is fairly well documented. David Marr did some very influential theoretical neuroscience papers where he proposed what computations the anatomy implied, and then later moved to computer vision where he became influential.

It is quite easy to make the case that eg LeCun's vision solutions were generalizations from audio processing to image processing. In practice, we can often use AI models to interpret neuroscience, but we can rarely use neuroscience to predict how to build an AI model.

I thought historically LeCun was inspired by Fukishima's neocognitron, but could be wrong. I agree in general that it is difficult to go from brain science to AI, but am very interested when this does happen.

Jeff Hawkins, "On Intelligence".