Cipollini, B. & Cottrell, G.
UC San Diego
Left-right asymmetries have been noted in tasks requiring the classification of many types visual stimuli, including Navon figures, spatial frequency gratings, and faces. The Double Filtering by Frequency (DFF) Model (Ivry & Robertson, 1998) has been used to implement computational models accounting for the human behavior in each type of study above. This PDP model uses hard-coded spatial filtering banks and hand-selected weights to implement the critical asymmetric “attentional selection” of frequencies. Hsiao, Shahbazi, & Cottrell (2008) implemented the Differential Encoding (DE) model, a simple feed-forward autoencoder with sparse, random connectivity, where the model left (LH) and right (RH) hemispheres have equal numbers of connections, but the LH contains more distant connections with a higher probability than the RH-as found in “patch” connectivity in BA22 (Galuske et al, 2000). The DE model was shown to account for one of these datasets. Here, we implement the DE model using parameters that follow the published patch asymmetry more closely and successfully apply the model to three of the four datasets mentioned above. We then show that the DE is actually filtering spatial frequencies in a manner found in humans, with the RH favoring LSF information and the LH favoring HSF information. Finally, we show that these connectivity differences may be due to maturational differences between the hemispheres, with the RH beginning maturation earlier in development, when visual acuity is low (see Howard and Reggia (2007) for a review). We implement a feed-forward autoencoder with sparse, random connectivity, which prunes its connections with the smallest weights during learning. We show that this network tends to prune more distant connections when trained on blurred (LSF) images vs. full-fidelity (HSF) images. With reasonable assumptions about inter-patch pruning, we show that this leads to the asymmetry in patch connectivity found by Galuske et al.