Computational Maps in the Visual Cortex
     Figure 6.5
MiikkulainenBednarChoeSirosh
Home    
About the Authors
Back Cover    
Table of Contents 
Sample Chapter 
Figures    
References    
Errata    
Demos     
Talks/Courses 
Software    
Credits    
Purchase online at:

springeronline.com
amazon.com

Click on the image to see a PDF version (for zooming in)

Fig. 6.5. Role of ON/OFF channels in processing various kinds of inputs. Each row shows a sample retinal activation, the LGN response, the V1 response and its histogram (OH) for the ON/OFF LISSOM network, and the V1 response and its histogram (RH) for the reduced LISSOM network. The sample inputs are plotted in gray scale from black to white (low to high) and the LGN activations by subtracting the OFF cell responses from the ON. In the V1 plots (c,e), orientation preferences of those neurons that respond are color coded according to the key on top, and color saturation represents the activation level (selectivity is not shown to match the perceived orientation measure; Section 7.2.1). The two networks respond similarly to an oriented Gaussian input on a blank background (top row), which is why very similar orientation maps developed in Figure 6.4. As seen in the histograms, only neurons with orientation preferences matching the input line respond. However, the networks behave very differently for other types of input. The ON and OFF channels filter out nonzero background levels and smooth, gradual changes in brightness, which ensures that V1 ON/OFF responds only to oriented patterns and sharp edges (second row; the response is strongest on the bright side of the edge because only bright Gaussians were used in training, as shown in the top row of Figure 5.13). In contrast, overall background illumination with no edges is ignored by the ON/OFF network (third row), whereas it activates nearly all of the V1 neurons in the reduced model. Without the LGN, the response to most patterns is determined by the total amount of brightness in the input, rather than by the orientation preference of the V1 neurons. Nonzero background levels, gradual changes in illumination, and large, bright objects are all common in natural images (bottom row), and thus the ON and OFF channels are crucial for preserving orientation selectivity when processing such images. On the other hand, the ON and OFF channels can be omitted for networks that process only schematic patterns on a blank background. The natural image is a retina-size detail (as shown in Figure 8.4e) from National Park Service (1995).