Affinity Spots
PhD student Dylan Paiton produced this plot for a recent paper submission. The model is a two-layer neural network trained on videos of natural scenes. Each pair of plots corresponds to a layer 2 neuron. The square plot is the spatial "receptive field" of the layer 2 neuron, the circular plot is the time-frequency "receptive field" of the layer 2 neuron. At the start of the model, each layer 2 neuron is connected randomly to all of the layer 1 neurons. Over time, after looking at millions of video segments, it learns a unique and interesting connectivity pattern to specific layer 1 neurons.