Friday 13th January 2017
Ground Floor Seminar Room
25 Howland Street, London, W1T 4JG
Navigation with recurrent neural networks: from 2D path planning to 3D path integration
Spatial navigation is an ethologically relevant behavior that requires memory-based route planning towards intended goal locations. Place cell activity in the hippocampus has been proposed as a potential neural substrate for goal-directed behavior and recent experimental evidence unveiled hippocampal neural correlates of navigation before or during the execution of a planned trajectory ultimately leading to a goal. I will describe an attractor based model of place cell activity, augmented with sustained negative feedback dynamics (e.g. short-term synaptic depression) that accounts for these observations and supports the hypothesis that recurrent dynamics in hippocampus subserve navigation-related computations. In the second part of the talk I will focus on a fundamental component of path integration theories, the representation of head-direction (HD). A recent study uncovered the existence of HD coding in 3D. We developed a neural network model that exhibits activity patterns mapped continuously to the (non-commutative) group of 3D rotations, a non-trivial extension of commutative 2D HD models. We took advantage of the generality of 3D rotations encoding to implement a simple model of recognition of 3D shapes, which accounts for the psychophysics observation of “mental rotations”.
Part 2 is joint work with Hervé Rouault and Alon Rubin