| Abstract |
Both the brain’s sensory systems and many neural network models in AI are organised topographically: They embody the layout of the sensor array, such as the retina or cochlea. I will discuss recent finding from my lab showing that this topographic processing is not just an artefact of the organization of the input, but constitutes a fundamental mechanism used throughout the brain.
We have recently shown that even in brain regions devoted to memory-related and motor-output processing - tasks considered very "high-level" - the brain’s responses are organised topographically. That is, we can find maps of “retinotopic" visual space in cerebellum, the default-mode network, and the hippocampus. Apparently, even our abstract, internally generated thought processes maintain some sensory structure.
I argue that for brains that are in constant interaction with the environment through the senses, this mode of processing is required: all processing, however abstract, has to refer back to concrete sensory events. Moreover, sensory-based processing provides the brain with a common language, a reference frame for all brain regions to communicate in.
These findings allow neuroscience to once again serve as inspiration for AI. More ‘general AI’ will need to be devoted to successful interaction with the world beyond mere classification of inputs, which requires creating models of what’s going on in the outside world. Sensory processing seems a fundamental ingredient in the brain’s model-building, and might provide a route to allow similar model-building by artificial means. |