Neural dynamics of visual speed discrimination and motion perception

A neural network model of visual motion perception and speed discrimination is presented. The model describes the functional properties of individual visual neurons and their interactions. Input to the model is related to visual stimuli and output to psychophysical results concerning human visual perception. This model extends in several ways the Motion Boundary Contour System (Motion BCS), a model of cortical magnocellular processing which has previously explained many data, including: short-range and long-range apparent motion, first and second order motion, and motion segmentation and grouping effects. The first extension is the incorporation of speed-sensitive cells into the model. The new speed-sensitive Motion BCS shows how speed-tuning can be derived using spatio-temporal filtering and competition over multiple spatial scales of motion-sensitive cells. Using a range of scales, each of which is sensitive to a different range of input speeds, the model reproduces empirically derived speed discrimination curves and simulates data showing how visual speed perception and discrimination can be affected by stimulus contrast, duration, density and spatial frequency. Incorporation of directionally-sensitive transient cells into this speed-sensitive model shows how a system using only orientationally-unselective transient cells can become sensitive to direction of motion. The model is then further extended to show how spatial pooling of the outputs of the directional and speed-sensitive cells can explain various more complex motion phenomena. Long-range filtering and feedback are used to track the motion of featural points through purely motion derived signals. The propagation of signals derived from the movement of these featural points explains how correct directions of motion can be observed in the absence of unambiguous local directional information, as in the center of a moving line or other more complex forms. In addition, the coherence and perceived motion of multiple moving lines (plaid patterns) is modelled in terms of the interactions between tracked featural points and local motion signals.