Here's what I'm working on this summer:



It's an implementation of a spiking neural network, using spike timing dependent plasticity (STDP) to learn.

The screen shot above is from the model application I built. Visualization is done in Java3D. Green nodes are excitatory neurons. Red nodes are inhibitory neurons. And blue nodes are delay neurons, which means they just output the activation of other nodes, delayed by a specific amount of time. The direction of connectivity is indicated by the colors of the endpoints of the connections. Connections go from (magenta) to (cyan).

This particular network is designed to learn temporal XOR, that's the XOR problem, but with the inputs presented as a sequence, rather than concurrently. The four green units at the bottom are the inputs. The blue units at each level are delays that should enable the network to associate "what just happened" with "what's happening right now" so that it can form conjunctions between consecutive parts of a sequence. The red units at each level serve as winner-take-all subnetworks, making sure that only one green neuron at that level is firing at a time, while suppressing the others. And the two green units at the top are the output units.

It's not functional yet, but the visualization is working, so I thought I'd share.

Meanwhile...I'm going to California to visit a friend for about a week, so blogging will probably be light to non-existent for a bit.