Institute for Sensory Research
Syracuse University
621 Skytop Road
Syracuse, NY 13244-5290
Phone: 315.443.9714 (lab)
Fax: 315.443.1184

 

Julian Martin Fernandez

 

Neural Model of SFM Computation

 

During my postdoctoral training at Dr. Ning Qian’s Lab (Columbia University) I proposed the first physiologically plausible model of SFM computation. It incorporates properties of neurons in visual areas MT and MST (Fernandez, Watson & Qian, 2002; Fernandez & Farell, 2005). A key assumption in the model is that the perception of depth from motion is related to the firing of a subset of MT neurons tuned to both velocity and disparity. The model MT neurons are connected to each other laterally to form modulatory interactions. The overall connectivity is such that when a zero-disparity velocity pattern is fed into the system, the most responsive neurons are not those tuned to zero disparity, but instead are those having preferred disparities consistent with the relief structure of the velocity pattern.

An important assumption that we incorporated into the model is the fact that humans can recover only relief structure from motion; i.e., an object’s 3D shape can only be determined up to a stretching transformation along the line of sight. This simplifies the computational problem posed to the visual system (and to modelers too!), and allowed us to build a successful model. Previous models of SFM computation intended to recover the full veridical shape. This is probably why previous attempts failed in implementing the models as physiologically based models and remained instead at the algorithmic stage.

Current work on the model concentrates in two main issues:

·      Learning mechanisms. In particular, I’m extending the model to incorporate self-organized learning, using LTP/LTD, to generate the modulatory connections between the model’s cortical areas MT and MST.

·      Extending the model to include arbitrary rotations. Our first version of the model, described above, only deals with the case where the axis of rotation lies in the frontoparallel plane. We are extending the model to include general rotations. This is being done in conjunction with additional experiments on SFM perception, so we can test the model predictions with actual psychophysical data. In particular, one of the problems is to understand how motion and stereo cues interact to generate the final percept, as both cues are intertwined and interact in various ways in the model (Fernandez & Farell, 2005e, coming soon).