Abstract
Problem: Motion detection responds to items within restricted regions in the visual field. Representations are facilitated by more global integration of motion responses to reduce initial stimulus uncertainty (Born & Bradley, Ann Rev Neurosci, 2005). Early stages of cortical processing of motion advance the generation of spatiotemporal input responses in area V1 to build feature representations of direction and speed in area MT. The neural mechanisms underlying such processes are not yet fully understood. Method: We propose a neural model of hierarchically organized areas V1, MT, and MSTl, with feedforward and feedback connections. Each area is formally represented by layers of model cortical columns composed of excitatory and inhibitory neurons with conductance-based activation dynamics. Receptive field sizes increase from V1 to MSTl as feedback kernels decrease from MSTl to V1 in a reverse hierarchy (Hochstein & Ahissar, Neuron, 2002). Top-down feedback and lateral connections enhance activations by modulatory interaction. Together with pool normalization this realizes a distributed up- and down-modulating gain control mechanism. Results and Conclusion: Stimuli motivated by psychophysical experiments were used to probe the model. Model MT feedback modulation and lateral competition leads to de-emphasized normal flow responses in V1 and enhanced intrinsic terminator motions. The speed of component motion of shapes is estimated by maximum likelihood decoding of MT population responses of speed selective cells. Aperture motion was disambiguated by activation growth-cones spreading from feature points via recurrent model MT-MSTl interaction to adapt the populations’ direction tuning generating coherent moving form representation. Its temporal signature depends on the boundary length (Pack & Born, Nature, 2001). Lesions of the model demonstrate that especially the feedback connection schemes were indispensable to feature enhancement and shifts in feature coding. Our results suggest how V1, MT, and MSTl feed-forward-feedback processing builds a coherent representation of moving forms using distributed representations of different spatiotemporal features.
Acknowledgement: Supported by Baden-Württemberg foundation (project no. NEU012).