Addressing the role of feedback

A critical part of the model is the use of feedback connections through which information from one modality influences lower processing stages of the other modality. In collaboration with Professor Geoff Hinton, I investigated another family of unsupervised learning algorithms that have a similar supervisory role for feedback connections. In these models, the feedback provides a generative model of the sensory environment. The feedback learns to predict the feedforward signal and simultaneously teaches the feedforward connections to represent predictable structure.

The models predict that feedback connections should not be plastic at the same time as feedforward connections. I investigated this property of feedback in a mouse slice preparation at UCSF. I found that feedback connections were not plastic under a protocol that induced plasticity in both forward and horizontal (lateral) connections.

Most recently with my graduate student Tom Sullivan, we have investigated the conundrum that feedback connections are almost exclusively excitatory (from excitatory neurons to excitatory neurons) but that the physiological effects of removing or disabling feedback is a removal of suppression from surrounding stimuli (and a small decrease in response to stimuli with no surrounding stimuli). In other words removing feedback seems to largely remove suppression as if feedback is providing a suppressive or inhibitory effect. Our model shows that this can be achieved with only excitatory feedback and that the suppression can be explained as a decrease in excitatory feedback due to competition at the higher level area. Our model accounts for several effects of surround suppression including: a change in receptive field size with contrast, surround facilitation at low contrasts and surround suppression at high contrasts, and different effects in different neurons with one consistent computational rule.