Figure 3.

The discriminability of responses estimated using a SOM neural net. The spatio-temporally patterned array of inputs is weighted based on unsupervised learning to cluster the sets of inputs according to the variability and the patterns present in the dataset. Large intrinsic differences in patterns between responses to 2 stimuli thus lead to reliable clustering. We compare this SOM decoder where each timed point and each neuron are weighted independently (magenta) to a decoder based on Euclidean distance with time points and neurons kept as separate dimensions or a van Rossum metric where responses of different neurons are averaged together before the comparison. (A) PC neuron responses to electrosensory chirps. (B) PN response of the moth antennal lobe to odors. (C) LIF model responses to frozen white noise stimuli of different shapes. Curves show averages (± s.d.) across all pairs of stimuli (number of stimuli: electrosensory = 3 different chirps, 3 pairs; olfactory = 7 odors, 21 pairs; LIF = 10 noise patterns, 45 pairs).