|
Towards understanding visual cortex and efficient data representations in deep learning| old_uid | 16855 |
|---|
| title | Towards understanding visual cortex and efficient data representations in deep learning |
|---|
| start_date | 2018/12/04 |
|---|
| schedule | 14h30-16h30 |
|---|
| online | no |
|---|
| summary | Recent advances in neuroscience have brought an unprecedented growth in the scale, structure and complexity of data to be processed. In this scenario it is crucial to understand the principles, constraints and goals of neural computation to extract meaningful information. The starting point of my work is the intuition that one of those principles allowing for efficient computation is symmetry. For example in the vision domain this corresponds to the ability to grasp the essential structural character of an object even if presented in a great variety of different aspects and viewing conditions (thought as symmetries). I propose the use of image symmetries, in the sense of equivalences under image transformations, as data dependent-priors for learning symmetry adapted representations, i.e., representations that are invariant to these transformations. I show how the visual cortex can implement such a representation drawing a parallel with architectures of Deep Convolutional Neural Networks. I end showing how the principle of invariance allows to predict specific neuronal tunings in V1, in the face patch of the monkey cortex and how it gives a possible explanation of the presence of specialized moduli in the cortex. |
|---|
| responsibles | Sarti, Petitot, Nadal |
|---|
| |
|