#self-organizing map

[jplotting]

   Self-Organizing map simulations of a 20x20 feature map. The data set is a simple set of 200 to 300 2D data points that have been randomly placed within several cluster spaces. Self-Organizing maps are good at data and dimensionality reduction in such that they take any n-dimensional data (like 2D vectors or 3D colors) and compress them down into a 2D dimensional representation (sometimes 1D or 3D). This representation is the top-right-most greyscale image called a U-Matrix.

  Learn More!

[autismscience]

Idiosyncratic perceptual transformations & atypical category learning in autism (self-organizing map simulations).

“Children with autism spectrum disorder (ASD) sometimes have difficulties learning categories. Past computational work suggests that such deficits may result from atypical representations in cortical maps. Here we use neural networks to show that idiosyncratic transformations of inputs can result in the formation of feature maps that impair category learning for some inputs, but not for other closely related inputs. These simulations suggest that large inter- and intra-individual variations in learning capacities shown by children with ASD across similar categorization tasks may similarly result from idiosyncratic perceptual encoding that is resistant to experience-dependent changes. If so, then both feedback- and exposure-based category learning should lead to heterogeneous, stimulus-dependent deficits in children with ASD.”