 |
Center for Nonlinear Studies |
Computations in neural circuits are inherently flexible, allowing humans and animals to respond to sensory stimuli with actions that are appropriate in a given context. Fundamental to this flexibility is the ability to integrate only context-relevant sensory information while ignoring irrelevant, distracting information. We studied the neural mechanisms underlying such context-dependent gating in monkeys performing two different sensory discriminations on the same set of visual stimuli. A contextual cue instructed the monkeys to report either the direction of motion or the color of a noisy visual stimulus. While the monkeys performed this task, we recorded neural responses from several cortical areas contributing to the monkeys' choices. We found that the gating of relevant sensory signals, and their integration towards a choice, can be understood as two aspects of a single dynamical process reflected in the responses of populations of pre-frontal cortical neurons. Using linear regression, we identified a multitude of signals represented simultaneously in the responses of pre-frontal neurons, including the direction of motion and the color of the stimulus, the context, and the developing choice. While these different signals are mixed at the level of single neurons, by projecting the neuronal activity onto the regression vectors, these signals are separable at the level of the population. In order to understand better the nature of the mixed signals in pre-frontal neurons, we trained a recurrent network model to solve a similar task, that of contextual integration of only one of two noisy input streams. The model reproduces the dynamic representations of the relevant signals, such as the motion and color stimulus, and the developing choice. Additionally, the model reveals previously unknown mechanisms for integrating the relevant input stream while ignoring the irrelevant one. We found that the network created two context-dependent, approximate line attractors to integrate the relevant sensory inputs. Surprisingly, the precise dynamics of integration are not simply based on a non-zero projection of the relevant input vector onto the relevant line attractor, but rather results from the transiently expanding, yet stable dynamics at each fixed point on the line attractor. Geometric reasoning suggests that this solution is highly likely given the observed orthogonality between the network input and output vectors.
Host: Garrett Kenyon, gkenyon@lanl.gov, 7-1900, IS & T