with the department heads of Brain and Cognitive Sciences (BCS) — Michale Fee and Josh McDermott.
We have a strip of our cortex that lights up for speech (even in a foreign language) that is totally separate from the strip that lights up for music (including ring tones and whistling). And none of the language centers light up during non-linguistic tasks (like math or music). Chomsky was wrong: language does not precede thought.
And musical preferences are not instinctual either. Our preference for fifths and octaves and our distaste for screeching high-frequency sounds (like nails on chalkboard) are not universal. In the jungles of Bolivia, the preferences are reversed. They hypothesize that common preference in urban areas comes conditioning in the presence of industrial noise.
“AI is the single biggest change in my field in the last 10 years.” When training a large AI model on auditory tasks, the results recapitulate the behavior and subsystems we see in the human auditory cortex. The behaviors and limitations of both brains and AIs come from a common optimization task. The AI model is a tool for understanding how much of leaning can be captured as an optimization. And it can help develop much better hearing aids and cochlear implants.
I asked about analogs to critical periods in childhood development and whether they would tune the hyperparameters of metaplasticity while training their models. He said that is their next frontier – leaning and plasticity. And looking beyond optimization to homeostasis in the model parameters.
Biggest bottleneck in their research? The amount of compute available.
Not many animals learn their vocalizations. No other primate does. Animals have no grammar, but the theory of dependency grammar suggests that grammar is just the connection between words. Like a LLM, there are no other hidden structures beyond those word pairings. With ML, we might be able to finally understand how animals communicate.
An aside: 50% of all MIT undergrads choose CS as their major.
Leave a Reply