Unsupervised detection of repetitive activity in neural populations

Matt Mahoney

Dept of Neuroscience, University of Vermont

Memory consolidation is understood to be one of the crucial functions of sleeping. Correlations between (lack of) sleep and cognitive performance are well-documented, but the mechanistic underpinnings are only just becoming clear at the level of single neurons. In the last decade, it was shown that certain cells in rat hippocampus which encode spatial information (so called place cells) are reactivated during sleep in temporal patterns that are similar to their patterns during survival-salient behavior while the rat is awake. For example, when the animal is trained to run along a track for a food reward, it encodes a sequence of place cell activity. This pattern is "replayed" when the rat is asleep, indicating that explicit repetition of the neural correlate of experience is part of the memory consolidation process. This phenomenon was discovered by a "template matching" analysis that searches in sleep data for an experimenter-derived expectation of the repeated pattern (the template). From a functional perspective, however, it is not clear what constitutes a salient memory. In other words, why would the rat replay the maze and forsake all other experiences? The answer is clearly that it doesn't, but finding such novel replayed patterns requires an unsupervised approach to the template matching method. I will discuss such an approach that uses the full windowed autocorrelation matrix, C, of a time series to mine for repeated patterns independent of the prejudice of the experimenter. The matrix C is dense but structured so it can be filled and manipulated efficiently for many classes of operations. In particular, it can be used to build a Markov chain whose stationary distribution can be used to identify the repeated patterns in the data. I will demonstrate the effectiveness of this approach on data with known patterns and, if time permits, I will discuss issues with scaling this technique to long recordings (106 time points).

Back to ACMS schedule