Biologically-plausible neural models for time-series clustering

The brain processes sensory data in an online fashion, incorporating information as it is being received. Typically this data does not come in independent samples, but follows some stochastic dynamics with parameters that are subject to change. For instance, the output of an olfactory sensory neuron might fluctuate around a mean that varies depending on the relative motion between the animal and an odor source. An important problem in this context is to find when a change in the dynamics occurs. In biological terms, this might correspond to a significant change in the environment that requires an animal to change its behavior. Here we build biologically-plausible algorithms for performing online clustering of time series data. These models can be implemented as neural networks with local learning rules, and can be derived from principled objective functions. We test these algorithms on data generated from several alternating autoregressive-moving-average (ARMA) models and find very good performance in detecting changes in the generating process within just dozens of samples from the transition point. We compare this to similar methods from control theory that are not directly interpretable as neural models.