Self-Organizing Temporally Coded Representation Learning

The self-organizing map (SOM) is an unsupervised learning algorithm that extracts representations from an input dataset and organizes them in a topographic manner. Nevertheless, the SOM is unable to handle event-based and asynchronous data such as spikes. This work introduces a spiking SOM that consists of a network of leaky integrate and fire neurons. Our spiking model differs from previous ones by demonstrating not only the ability to generate topographically ordered maps, but also the additional capability of vector quantization (VQ). Thus our model replicates for the first time the two key functions of SOM. To do so, we extend the VQ capabilities of a previous model by incorporating a novel neuromodulator, which enables the generation of ordered maps. We demontrate good performances on synthetic and real datasets.