A Single Learning Rule can Account for the Development of Simple and Complex Cells M Teichmann, F Hamker |
|---|
Chemnitz University of Technology, Germany
|
| Understanding the human visual system and the underlying learning mechanisms is a vital need for computational models of perception. One open question for the development of such models is how the visual system achieves its ability to recognize objects invariant to various transformations. To study potential mechanisms for learning this invariant processing, we created a multi-layer model of the primary visual cortex (V1). The model consists of an input layer simulating the LGN input into V1, the so-called simple-layer related to V1-layer 4, and the complex-layer related to V1-layer 2/3. In our previous work [Teichmann et al, 2012, Neural Computation, 24(5), 1271-96], we found that trace learning is a suitable mechanism for learning the responses of V1 complex cells. Here we show that a single learning rule can account for the development of simple- as well as complex-cell properties. We apply this learning rule to exploit the temporal continuity of the visual input, using a short-term trace for neurons in the simple layer and a longer-term trace in the complex layer. We show that neurons in the simple layer develop receptive fields comparable to monkey data, while neurons in the complex layer exhibit phase invariance. |
| Up Home |
|---|