Learning Features of Simple and Complex Cells: A Generative Approach via Multiplicative Interactions

Wentao Huang, Wentao Huang, Zhengping Ji, Garrett Kenyon
2011 Nature Precedings  
Goal: •A computational model to learn the feature bases (receptive fields) of simple and complex cells in the primary visual cortex. •Address the translation invariance developed from c cells via natural image sequences. Approach: •Reconstruct an input via a linear combination of feature bases in simple cells. •Modulate the simple cell representation via multiplicative interactions from complex cells. •Enforce a sparseness prior for the latent representation of simple cells and complex cells.
more » ... nforce a slowness prior and a trace-like rule for the representation of complex cells. Advantages: •Provided a factorized approach via the product of twoorder tensor weight parameters and only one latent variable for invariant representation, more efficient than bilinear models that contain three-order tensor weight parameters and two latent variables. •Demonstrated general simple cell feature maps and complex cell invariant receptive fields simultaneously.
doi:10.1038/npre.2011.5943.1 fatcat:v2axlfbhjjfrbpisr7jx64hmbi