Fig. 1

DRREP Architecture DRREP is composed of N Sub_DRREPs. Each Sub_DRREP’s first layer, L1, is composed of S neurons, each of which is made out of a randomly generated k-mer based mismatch function. The second layer, L2, is composed of S nodes, with the entire layer performing normalization of the L1’s signals. Layer L3, is the learning linear neural layer, whose synaptic weights are calculated using the Moore-Penrose generalized inverse. All N Sub_DRREPs are stacked in parallel. The L4 is a norm-pooling layer, composed of N nodes, which normalizes the signals from each Sub_DRREP. The next layer, L5, is composed of a single thresholding neuron, which weighs each contribution from the Sub_DRREPs based on that Sub_DRREP’s relative validation score, and passes this value through the threshold to output the final score for the input sliding window