Abstract
The visual system constantly adapts to the statistical structure of its inputs. Recent experiments suggest that surprisingly adaptation in V1 can maintain population homeostasis, e.g. roughly equal responses of cells across preferences despite biased stimuli (Benucci et al., Nature Neuroscience, 2013). However, it remains open how the system achieves such sophisticated gain control. To better elucidate the computational mechanisms, we formulated a network model for V1 adaptation. The model neurons receive inputs via both feedforward and recurrent connections. We incorporated multiple neurophysiologically realistic mechanisms for adaptation, including short-term plasticity (e.g., pre-synaptic depression), as well as adaptation currents proportional to the neuron's integrated firing history exponentially weighted at multiple time scales. We simulate the neural responses based on both uniform and biased stimulus ensembles. For the latter, one particular orientation (the adaptor) is over-represented. We systematically vary the strength of pre-synaptic depression and the adaptation current, and find that, for a range of parameter values, the resulting network can well maintain first-order (mean responses) and second-order (correlations) homeostasis, consistent with Benucci et al. Also, the model robustly accounts for the experimentally observed repulsive shift of tuning curves and population activity during adaptation. Furthermore, fitting the model neural responses using a two-gain-factor model (stimulus-specific and neural-specific gain) proposed by Benucci et al., the resulting gain factors are comparable to the experimental report, with the stimulus-specific gain generally dominating. Our model also generates new predictions, including that the amount of response equalization depends on the contrast. In particular, stimuli with lower contrast should lead to less perfect equalization. Our results suggest specific mechanisms that could maintain population homeostasis during a particular set of adaptation experiments. In general, our model may provide a modelling framework for studying other commonly used adaptation protocols in psychophysical and neurophysiological experiments.
Meeting abstract presented at VSS 2018