4 ms·
> Another salient property of the brain ... is that the connection strengths between neurons can be modified in response to activity and experience—a process th
by saeranv 4y ago
> Another salient property of the brain ... is that the connection strengths between neurons can be modified in response to activity and experience—a process that is widely believed by neuroscientists to be the basis for learning and memory. Repetitive training enables the neuronal circuits to become better configured for the tasks being performed, resulting in greatly improved speed and precision.
This seems like a clear description of a bayesian process. The new connections are a posterior updated from the likelihood informed from input sensory data. But the fact that they don't describe it as such makes me wonder if there's other reasonable ways to model this learning behaviour.
I am familiar with Karl Friston's work (as a layperson) of animal behaviour and learning as a form of bayesian update via variational inference. Does anyone know if his ideas, or the general idea of learning as a bayesian process well accepted in the neuroscience industry?
- EricMausler 4y agoNeurons that fire together, wire together
- eli_gottlieb 4y agoThe general idea of a Bayesian brain is fairly well-accepted. Not everyone agrees that oh, this must be the way the brain works, but there are definitely experiments demonstrating set-ups in which you can get a good approximation of a Bayesian posterior distribution out of subjects. The question is how the brain might do Bayesian inference in a very broad class of generative models (not just affine-Gaussian!), and what else it might be doing inference in service to.
- saeranv 4y agoThat makes sense. Since you seem to have expertise in this domain, may I ask what is the current state-of-the-art for generative models for modeling learning/behaviour, in the neuroscience community? If any...
- eli_gottlieb 4y agoThe neuroscience community doesn't really keep "benchmarks" the way the machine learning community does. The trouble this causes for us right now is that many experimental tasks now used in neuroscience are simpler than the corresponding tasks used in ML, and so could be "solved" by a significantly simpler model. That said, I've seen stuff like the Hierarchical Gaussian Filter used to fit behavioral data: https://link.springer.com/chapter/10.1007/978-3-030-93736-2_57 https://link.springer.com/chapter/10.1007/978-3-030-93736-2_...