Orientation: Initial

Spinning line and cortical response Orientation key Orientation map
a b c d e
f
g h
Key:
  1. retinal image presented to map g.
  2. initial response of the cortex to that image
  3. orientation histogram of the initial response
  4. settled response of the cortex to that image
  5. orientation histogram of the settled response
  6. key showing color assigned to each angle
  7. self-organized RF-LISSOM orientation map
  8. orientation histogram of the orientation map
The sequence repeats after the image has made the full circle. A slower-loading version of this page with larger pictures is also available.

These images illustrate how an RF-LISSOM network responds to oriented lines before self-organization. It's easiest to understand them if you first start with the self-organized version. Box (g) shows an orientation map measured for an RF-LISSOM network before self-organization; each neuron is represented by the color of the orientation it prefers. The colors for each orientation are shown in the key beside it. At this point, the orientation preference is random for each neuron, which is why the colors look like static.

The animated panels show the response of this simulated cortex to a single blurry line at the center of the retina, using the same color code for each neuron. The rotating grey line (a) shows the input presented to the retina. The fuzzy colored areas (b) show which neurons responded at the instant the activation reached the cortex. For this simulation, the neurons have been initialized to respond to the location on the retina corresponding to their location in the cortex, which is why the activity shows up in the center of the cortex for this input in the center of the retina. The histogram adjacent to the initial response (c) shows the total amount of activation in the initial response for orientation detectors of each type. The sharply defined areas (d) show the neurons still responding after lateral interactions have taken place. Finally, the second histogram (e) shows the total amount of activation in the settled response, for each type of orientation detector.

Clearly, this network is not encoding orientation: the histogram of activated units remains almost entirely flat regardless of the input. Moreover, the network is to a large extent not even preserving orientation. Due to settling by the lateral interactions, the activation pattern barely differs at all as the input is varied, remaining a nearly circular blob. This suggests one reason why orientation preference is such a prominent feature of the earliest cortical areas; if it weren't encoded there, orientation information would be lost forever.