Talk Summary: The speaker presented a model of V1 that produces a saliency mapl. This model transforms contrast inputs to saliencies using contextual influences. The contextual influences are implemented in a recurrent network with intra-cortical connections. Saliencies are signalled using firing rates of the neurons.The speaker compared the performance of her model with known psyho-physical results.
Discussion And Questions Following the Talk
Q(Bill): Do all V1 cells signal saliency or a few of them.
A: Yes all cells.
Q: What is the role of other visual areas in this model?
A: No role.
Speakers' Comment: V1 cells firing rates signal saliencies, despite their feature tuning. Strongest response to any visual location signals saliency. Also, this theory is only bottom up. V1's output as saliency map is viewed under the idealziation of the top-down feedback to V1 being disabled. eg. shortly after visual exposure or under anesthesia. (Saliency from bottom up factors only)
Q: Are these results after the system have settled?
A: After one time constant.
Q: Long range excitatory connections are not isotropic. In your figure it seems as if they are not.
Comment: The speaker then explained how they are not isotropic in her model.
Bill: What about color, motion etc. how do u combine in the model?
Ans : These are not included in the current model.
Q: Did you do any of these on natural images?
A: No the limitation is the processing power.
Q: Do you need any endstopping?
A:
Steve (Comment): I am worried about about the use of the word "saliency" map for the visual search tasks you mentioned. For eg we have found that 3D properties (surfaces) can affect pop-out mechanisms and the word 'saliency' is usually used in a more wider context. Have you done any experiments using surface properties?
Answer: I just compared with visual search tasks. This model does not explaining 3D surface properties.
Dana(Comment): The same data can be explained using Signal to Noise Ratio. (Triesmann data) Some of the experimental data could be false because they were done in blocks.
Q:What about multiple scales?
A: Same principle will apply at all scales.
Q: Regarding the disynaptic inhibitory connection, is there a time delay that is required?
A: Yes. To get rid of the symmetry breaking behaviour we need the disynaptic behaviour.
Back to the
main index
for Inference and Prediction in Neocortical Circuits.