|
Understanding neural computation from the perspective of the brain as a learning machine| old_uid | 7627 |
|---|
| title | Understanding neural computation from the perspective of the brain as a learning machine |
|---|
| start_date | 2009/11/17 |
|---|
| schedule | 11h |
|---|
| online | no |
|---|
| location_info | salle de conférence |
|---|
| summary | One may argue that it was relatively easy for evolution to create networks of neurons with large computational power, since in principle every computation can be carried out on rather simple feedforward neural networks, and the structure of the network is not even critical for that. In contrast, it was rather difficult for evolution to create brains that can learn autonomously, and can maintain stable computational function in spite of noise and various types of ongoing changes in the system. However, research results from machine learning have shown that cleverly designed circuitry can drastically improve the learning capability of an autonomous learning system. Hence I propose to understand generic cortical microcircuits, and possibly also the large scale connectivity structure of the brain, from the perspective that these architectures provide a clever scaffold for learning, in fact for a variety of learning mechanisms that act on different temporal and spatial scales
I will illustrate this hypothesis by discussing a new result on learning in Winner-Take-All (WTA) circuits of neurons, a common network motif of cortical microcircuits. It turns out that STDP (spike-timing-dependent plasticity) induces and maintains a powerful computation role of such circuits: It enables such circuits to create a sparse encoding of complex high-dimensional spike inputs. In fact, one can prove that STDP approximates in this context Expectation Maximization, a powerful tool for discovering hidden sources of complex inputs. This is apparently the first link between STDP, the most commonly reported rule for synaptic plasticity, and the world of probabilities and probabilistic inference, which has turned out to be very suitable to describe higher level computational operations in the brain. |
|---|
| responsibles | Vezoli |
|---|
| |
|