Direction Map Demo

James A. Bednar and Risto Miikkulainen

This page shows results for a laterally connected self-organized map of orientation and direction preferences. If the preceding sentence sounds like gibberish to you, you might want to check out my research page and some of the other more heavily documented demos first. Even if it makes sense, you may need some aspirin after looking at this page long enough :-).

When trained with moving, oriented patterns, the HLISSOM model develops spatiotemporal receptive fields forming a global map of orientation and direction preference:

Animated direction/orientation map

Orientation key

(A version of this page with smaller pictures is also available. )

This map shows an array of 20,000 neurons. The color of each neuron represents its orientation preference, using the color key shown below the map. For instance, the neurons in the bottom left corner prefer nearly horizontal lines, and hence are colored orange.

The animated black lines superimposed on the orientation map are plots of the spatiotemporal receptive field of every fourth neuron. The animation has four steps, and repeats after the fourth step. The receptive field represents the stimulus that would be most effective at activating the neuron. For instance, the neurons in the lower left corner will prefer a nearly horizontal line, moving in a downward direction. These receptive fields are similar to those measured experimentally in animals (DeAngelis et al 1995).

As also found in animals, many of the patches of neurons selective for a given orientation are subdivided into patches preferring opposite directions of motion (Weliky et al 1996, Shmuel and Grinvald 1996). For instance, the largest blue patch (just below and to the left of center) is divided in half vertically. Those in the left half prefer motion down and to the right, while those in the right prefer the opposite direction. Similarly, the longest contiguous green patch of neurons (slightly above the blue one just mentioned) is subdivided into several different patches, each preferring a direction opposite to its green neighbors. Similar paired patches can be seen throughout the map, although there is sometimes a short distance between the patches preferring each direction.

These results show that a simple self-organizing algorithm can explain the development of the temporal preferences of neurons as well as their global arrangement into direction and orientation maps. For more information, check out our CNS*2002 paper.