Abstract
Our neurocomputational model suggests that a difference in connectivity between cortical patches in the left and right hemispheres (LH and RH) can drive lateralization in visual processing (Hsiao et al., 2013). Anatomical measurements of BA 22 show an asymmetry in the average distance of long-range intrinsic axons that interconnect groups of selectively interconnected neurons ("patches"). Our model consists of two autoencoders (neural networks trained to reproduce their input on their output), where the hidden units represent the interaction between the patches. Similar to the neural system, our hidden units selectively interconnect nearby input units, have no asymmetry in the number of interconnections, but are asymmetric in the average distance between interconnected inputs. The model reproduces a number of behavioral asymmetries found in the visual processing literature. We extract hidden unit encodings for task-specific stimuli, classify the encodings according to the task, and compare performance across LH and RH networks. In addition, our model spontaneously shows predicted asymmetry in spatial frequency encoding. Here, we address a potential critique of the model. Previously, our autoencoders were trained only on the task-specific images used in the behavioral experiments. Here we train the models with natural images, then present them with the task stimuli, and they still reproduce the behavioral asymmetries modeled previously. We also made the model more biologically plausible by adding homeostatic scaling of synaptic weights and denoising properties; this further enhances spatial frequency differences. Finally, we show that the LH model in fact exhibits no spatial frequency preference--it is the RH model that shows a bias for low frequency information, at the expense of high. This is further evidence that our model's behavioral asymmetry is driven by the anatomical structure of the model, and not due to other factors, such as an interaction between the training stimuli and the model architecture.
Meeting abstract presented at VSS 2014