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

Fig. 5.13. Effect of training patterns on orientation maps. In this and later similar figures, the rows represent different self-organization experiments. Each row typically shows a sample retinal activation, the LGN response to that activation, final receptive fields (ON-OFF) of sample neurons, their lateral inhibitory connections (LIs), the orientation preference and selectivity map, the orientation preference histogram, and the fast Fourier transform (FFT) of the orientation preferences. The RFs and LIs are drawn to a smaller scale than LGN and V1. For clarity, most OR models are based on abstract input patterns like the oriented Gaussians in the top two rows. However, OR maps develop robustly with a wide variety of input patterns, including large circular patterns (middle rows) and natural images (second row from the bottom; image from a dataset by Shouval et al. 1996, 1997). Maps develop even with random noise (bottom row), although such maps are relatively unselective and the RFs do not have realistic profiles. Spatial structure is therefore necessary in LISSOM for biologically realistic maps to form.