This demo shows how selectivity to intermediate complexity features (that may correspond to IT neurons) can emerge by repeatedly applying Spike Timing Dependent Plasticity on spike trains coming from V1 orientation selective complex cells. At the bottom of the screen we reconstructed the preferred stimuli of three IT neurons. As we start with random weight matrices, responses are chaotic at the beginning of the demo, but soon the three neurons detect statistical regularities in the input spike trains and self-organize, each one developing selectivity to a distinct feature.
Another similar demo with motorbikes.
A last one with a mix of faces, motorbikes, and backround images. Notice that 2 neurons developed selectivity to face features, and one to a motorbike feature. The background was too changing to be learned.
A more recent work, with a much deeper network with 3 trainable layers. The first layer learned edges, the second one intermediate complexity features, and the third one objects. Again, the background was too changing to be learned.
Masquelier T, Thorpe SJ (2007) Unsupervised Learning of Visual Features through Spike Timing Dependent Plasticity. PLoS Comput Biol 3(2): e31 doi:10.1371/journal.pcbi.0030031
Kheradpisheh SR, Ganjtabesh M, Thorpe SJ, Masquelier T (2016) STDP-based spiking deep neural networks for object recognition arXiv 1611.01421
Last updated Dec 13 2016