|
How and Why?| title | How and Why? |
|---|
| start_date | 2023/12/14 |
|---|
| schedule | 10h |
|---|
| online | no |
|---|
| location_info | Salle Jaurès |
|---|
| summary | Understanding specific brain regions therefore requires understanding how they perform their specific computations. Until recently we did not have plausible computational models of many of these functions. But all that has suddenly changed with the explosion of recent successes with artificial neural networks. In this lecture, I sketch the manyfold implications of these advances in AI for the organization and function of the human brain. First, artificial neural networks (ANNs) now succeed at many tasks similar to those conducted in specialized brain regions, from face recognition to speech perception and language processing. These ANNs thus provide the first computationally precise hypotheses for how these functions might work in the brain. Further, to a remarkable degree, responses in visual, auditory, and language cortex are well predicted by ANNs optimized for visual, auditory, and language tasks, respectively. For example, we can now predict with astonishing accuracy exactly how strongly the FFA, PPA, and EBA will respond to a novel image. But beyond providing testable and computationally explicit hypotheses for the computations conducted in the brain, these ANN models can inform not just how the brain works but why it works the way it does. With Katharina Dobs, we found that that a network trained on both face and object recognition spontaneously segregated itself into two separate systems, without any built-in priors to do so, suggesting that the statistics of experience may suffice, without face-specific innate predispositions, for the brain to achieve the functional organization it does. The stunning convergence between the organization and function of the brain and completely different and nonbiological ANNs optimized for similar tasks, is transforming cognitive science and neuroscience, from a focus on describing phenomena of the mind and brain and their underlying mechanisms, to a deeply theoretical enterprise of asking (and sometimes even answering) why they work the way they do. |
|---|
| responsibles | de Vignemont |
|---|
Workflow history| from state (1) | to state | comment | date |
|---|
| submitted | published | | 2023/09/26 14:05 UTC |
| |
|