|
Cerebellar learning with perturbations| old_uid | 17382 |
|---|
| title | Cerebellar learning with perturbations |
|---|
| start_date | 2019/02/18 |
|---|
| schedule | 11h-12h |
|---|
| online | no |
|---|
| details | Invité par Alberto Bacci |
|---|
| summary | A very general problem in learning is how to optimise a complex neural network when only crude sensory feedback from evaluation of behaviour is available. Minsky identified this issue and termed it the "credit assignment problem"; this is the problem that must be solved in "reinforcement learning". Minsky also suggested a plausible method for the optimisation, "stochastic gradient descent", but to date no complete implementation has been proposed for any brain region. We identify a credit assignment problem in cerebellar motor learning and propose a complete and biologically plausible cellular-level implementation of stochastic gradient descent. A central feature of our proposal is that the evaluation signal---climbing fibres reporting movement errors---can also serve to perturb the system spontaneously in a form of trial-and-error learning. Our theory predicts synaptic plasticity rules that appear to contradict the literature. However, by performing plasticity experiments under physiological conditions we support our predictions and show that the in vitro plasticity experiments are prone to generating erroneous activity-plasticity mappings. Finally, we demonstrate the theoretical capacities of our algorithm, which exhibits intriguing analogies with the learning of action sequences by the basal ganglia. |
|---|
| responsibles | Oliviero |
|---|
| |
|