next up previous contents
Next: 4.5 Orientation encoding Up: 4 Training the Orientation Previous: 4.3 Receptive fields and

4.4 Self-organization of lateral connections

The lateral connection weights self-organize at the same time as the orientation map forms. Initially, the connections are spread over long distances and connect to neurons of all types. As the lateral weights self-organize, the connections between uncorrelated regions grow weaker, and after pruning, only the strongest connections remain (figures 4.4 and 4.5).

 
Figure 4.4: Self-organization of lateral connections to a highly selective neuron.
  The plots show each type of weights going to the neuron at (88,111), both before (top row) and after self-organization (bottom row, after 30,000 iterations). A border is drawn around the active weight area in each plot, and the neuron itself is marked with a tiny box in (c). Bright colors at a location on the cortex in (b) and (c) indicate a strong lateral connection from that neuron to neuron (88,111); the hue encodes the orientation preference of the neuron. Initially, each neuron had lateral inhibitory connections to every surrounding neuron, so each connection was very weak. After organization, the neuron developed a preference of -30° (blue), and its connections come primarily from other blue neurons. The connections follow the twists and turns of each blue iso-orientation column (compare to figure 4.3b), and are elongated upon the orientation they encode.
\begin{figure} \begin{minipage} {\textwidth} \centering \begin{minipage} {\thr... ...\makebox[0pt][c]{ \usebox {\orkey} } \end{minipage} \end{minipage} \end{figure}


 
Figure 4.5: Self-organization of lateral connections to an unselective neuron.
  The plots show each type of weights going to the neuron at (88,105), both before (top row) and after self-organization (bottom row, after 30,000 iterations). The weights are colored as in figure 4.4. After self-organization, this neuron is in the center of a pinwheel, around which orientation preference changes continuously. The neuron is colored yellow because it has a very slight preference for +30°, but neighboring neurons of all orientation preferences connect to it.
\begin{figure} \begin{minipage} {\textwidth} \centering \begin{minipage} {\thr... ...\makebox[0pt][c]{ \usebox {\orkey} } \end{minipage} \end{minipage} \end{figure}

The strongest lateral connections of highly-tuned cells (figure 4.4c) link areas of similar orientation preference, and avoid neurons with the orthogonal orientation preference, as found in the cortex. Other neurons remain unselective (for instance, at points surrounded by continuously changing orientation preferences), and they connect to cells of all orientations equally (figure 4.5b,c). The connections of unselective neurons have not yet been studied, so this result represents a prediction of the RF-LISSOM model.

Furthermore, the connection patterns of highly oriented neurons are typically elongated along the direction in the map that corresponds to the neuron's preferred stimulus orientation (as verified subsequently for monkey cortex; Fitzpatrick et al. 1994). This organization reflects the activity correlations caused by the elongated Gaussian input pattern: such a stimulus activates primarily those neurons that are tuned to the same orientation as the stimulus, and located along its length (see Sirosh et al. 1996 for details).

next up previous contents
Next: 4.5 Orientation encoding Up: 4 Training the Orientation Previous: 4.3 Receptive fields and
James A. Bednar
9/19/1997