G. Major*; R. Baker; O. Debowy; E. Aksay; H.S. Seung; D.W. Tank
Recurrent network models of the oculomotor velocity-to-position integrator (VPNI) require a learning mechanism to tune synaptic feedback; mistuning results in leak or instability. Motivated by the idea that retinal slip might control learning, goldfish were presented with a visual pattern rotating with angular velocity proportional to horizontal eye position. When induced retinal slip was consistent with a leaky integrator, both horizontal eye position (>15 fish) and firing rates of position neurons in Area 1 of the medulla (6 cells) gradually developed instability: they diverged from a null point during fixations in the dark. The phase of low-frequency VOR was also delayed, implying changes were produced by an altered VPNI time constant. Conversely, when retinal slip was consistent with an unstable integrator, both neural firing rates (6 cells) and eye position (>15 fish) developed leaky integrator dynamics: they converged towards a null point and VOR phase was advanced. With several hours of training, the apparent VPNI time constant could be altered up to ten fold, from a typical control value of ~30 s (slight leak) to <5 s for instability, and to <3 s for leak. A stationary visual pattern following training accelerated recovery to control values. Our results suggest that plasticity normally tunes the VPNI by minimizing retinal slip. The experimental demonstration of such long hypothesized plasticity provides support for network models of the neural integrator that require precise tuning of recurrent feedback.
Supported by: Lucent Tech., NSF, NIH, Wellcome Trust and MIT