Several more days of neuromorphic fun. There are lots of things we don't understand about the brain: very little of the brain is active at one time, neurons send other neurons encapsulated mRNA messages, split brain patients become functions extremely quickly, dendrite processing is an important process. I led a session on embodied physical reservoir computing. A session on how LLMs work was illuminating. There is work on building biotransistors which can be made on flexible substrates, and work in water. The brain includes a "neuromorphic twin" model of the world, which it uses to predict sensory data; if there is a mismatch, there can be peculiar sensations. The brain has as many glial cells, including astrocytes, as neurons; a model of these at one extreme of a spectrum of structure looks a bit like a dense associative memory, what capabilities might other places on the spectrum provide? AI basically hasn't incorporated any discoveries from neuroscience in the last 50 years. Deep learning, with its billions of parameters, works because a particular Riemannian metric is more likely to exist in these high dimensions giving a unique global minimum. Insects diverged from other animals before there were brains, so have rather different structures. Bee navigation establishes an absolute reference frame by using the sun and time of day.
Whew! That's been a lot. I'm flying home tomorrow, so today I spent time walking, decompressing by watching the sea up close (it's profoundly weird to stand close to the shore for quite a while, and the tide neither comes in nor goes out), and a bit of paddling.
 |
| There's a lovely sandy beach in front of the hotel, which ends at these rocks. Standing listening to the water gently lapping against them is incredibly relaxing. |
No comments:
Post a Comment