Fig. 2.

Neuromorphic self-organized mapping and spiking classifier model. a Previous conceptual neuromorphic algorithm for a generic multivariate classifier design based on the insect olfactory system. Inputs were clustered using a (unsupervised) neural gas algorithm on a standard PC up front leading to VR${}_1$ to VR${}_N$. b Adapted conceptual model using the neuromorphic self-organized mapping (unsupervised learning stage). The previously used clustering of the input space to produce a set of virtual receptor points is replaced with the self-organizing set of RN neurons that are tuned to represent prototypes in the input space during a first pass through the training set. c The new model optimized for size whereby the RN and PN layers are merged into a single self inhibiting layer, now labeled CN (cluster neurons). d The final implemented model optimized further: Layers are now implemented as single populations with subpopulations demarcated by connectivity and the biologically correct inhibitory interneuron populations are made redundant by using direct inhibitory synapses between subpopulations. An additional population of Poisson neurons was added to implement supervised learning during the second phase of training, by exciting or inhibiting designated output subpopulations depending on the class of the presented input (“teaching signal”). The number of output neurons reflects the number of classes, e.g., 10 for the MNIST dataset; only 2 are shown in the figure for simplicity