[This was posted after the fact for record keeping purposes]
Hi all,
I hope everyone had a good 4th of July!
We’ll be meeting again this Thursday (7/11) at 5:30 PM in the Brochstein Pavilion. This week, we’re going to talk about the inscrutability of our current methods of artificial intelligence. One of the interesting things about the neural networks that are used in all forms of data analysis today is that they can’t be interpreted well (or at all). Computer scientists can train them to accurately determine whether a YouTube video has a cat in it, but how the cat is encoded in the network is a mystery even though we know the properties of every simulated neuron. This mystery might even be logically impossible to solve, raising some questions on the limits of being able to understand our own brains.
One of our members suggested the paper below as an addendum to our talk. In it, the authors used deep neural networks to create neural stimuli that pushed primate visual neurons well beyond activity levels seen before. This brings up some interesting ideas on receptive fields (e.g., are neurons traditionally associated with linear receptive fields actually optimized for line perception, or have we simply boxed them into that category?)
https://science.sciencemag.org/content/364/6439/eaav9436