We’re having our next Brains@Bay meetup next week, Wednesday April 13th at 10am PDT. The topic is exploring neuromodulators and how they might impact AI, from a neuroscience and computational perspective. We have a great lineup of speakers, with Srikanth Ramaswamy (Newcastle University), Jie Mei (The Brain and Mind Institute) and Thomas Miconi (ML Collective), followed by a discussion panel and Q&A.
If you can’t join live, we’ll share the recording after the event. You can find the details of the talks and RSVP here:
I think neuromodulators implement some sort of supervision and reinforcement, while core unsupervised learning works though neurotransmitters, at much finer grain?
If you are limiting the system under investigation to the brain then reinforcement learning and supervision can be understood to involve feedback from the environment. From that perpsective modulation is internal to the system so it would not be providing supervision or reinforcement.
You could choose a smaller system under investigation e.g. a neuron, then you could say the environment of the neuron is providing supervision and/or reinforcement and the environment is inside the brain. In that case the mechanisms for reinforcement could include things like spike timing, inhibitory neurons, neurotransmitters, and neuromodulation.
Modulation of the brain’s neural network has diverse scales in space and time. You could consider the modulation of a synaptic compartment. You could consider the social modulation of an individual’s desire to question or accept state propaganda.
Yes, I prefer finer grain analysis, esp. on the level of neurons. But the basic principle that applies across all scales is that, relatively speaking, supervision (an add-on) must be more coarse than core unsupervised learning. Primarily it’s more coarse in time scale, hence my distinction between neurotransmission and neuromodulation.