Spatiotemporal Pattern Detection in Multi-cell Recordings Using Unsupervised Learning
Abstract
Detection of spatiotemporal patterns have many applications in areas such as computer vision and data mining. Specifically, the analysis and mining of biological data with high dimensionality (e.g. multi-cell recordings, fMRI) are heavily dependent on detection of these patterns. In this thesis, we propose two unsupervised learning algorithms for obtaining filters that capture temporal patterns. In particular, we are interested in applying our methods for detection of regularities in multi-cell recordings of neurons. We propose two approaches: convolutional restricted Boltzmann machine (RBM) and convolutional denoising auto-encoder. The experimental results demonstrate that the proposed methods are able to detect temporal patterns in artificial data and multi-cell recordings from rat's brain. Moreover, we propose a Monte Carlo method for quantitatively evaluating the convolutional RBM by estimating the log-likelihood of data under the model distribution. The experimental results on test dataset of handwritten digits (MNIST), demonstrate that the convolutional RBM can learn a good generative model with small number of parameters.