|
Efficient codes for unreliable bodies, toward an information theory for embodied AI and robotics| title | Efficient codes for unreliable bodies, toward an information theory for embodied AI and robotics |
|---|
| start_date | 2024/04/11 |
|---|
| schedule | 14h-15h30 |
|---|
| online | no |
|---|
| location_info | Salle Marey |
|---|
| summary | I will present some recent findings and a proposal on yet another theory of embodied intelligence, called informational embodiment.
Taken individually, our muscles are limited in terms of force and precision, and our neurons are limited in terms of computational power. However, taken globally, our body and brain perform better than machines and AI systems, with less resources.
One principle enounced by Barlow (1969) is that biological systems do maximize information with respect to their computational material (sensors, muscles, neurons), and use efficient codes to produce optimal sensing, perceiving, learning and acting, despite the unreliability of the materials.
I will show how this principle may provide some guidelines to robotics and AI, to maximize information (entropy) and to remove the useless ones (redundancy) for smooth control, 3D perception, planning, and how it can make sense that soft robotics are maybe better on the long range than current robotics. |
|---|
| responsibles | Marin |
|---|
Workflow history| from state (1) | to state | comment | date |
|---|
| submitted | published | | 2024/04/03 13:47 UTC |
| |
|