Sorry about the late update but I got wrapped up in poster work. I lost the program on which I had written notes for Sunday, so I’ll be writing this off the top of my head. There was a series of posters from the Movshon lab showing some efforts towards undertanding V2.
One was a continuation of the work on receptive field models of V2. Last year the same subject was tackled and Tony confided that he was disappointed to see little difference between V2 and V1 receptive fields. This year was more conclusive, showing the presence of discrete nonlinear subunits in V2 cells not visible in V1. From a modeling persective, the most interesting aspect I thought was the use of a convolutional neural-net-type to look at tiled subunit structures. My current best try at the Neural Prediction Challenge in V1 is based on this very principle, and of course state-of-the-art V4 modeling is based on this as well. As a student from the Gallant lab told me (sorry, can’t remember the name), well nobody has any better ideas anyways!
The other two show kind of an interesting turnabout in Tony’s world philosophy. The man that hated natural images has started using natural textures to look at population coding in V2. Two results stood out:
- V2 is much more sensitive than V1 to natural textures
- V2 encodes natural textures along perceptual lines
Posters are available from the above links.