Learning and Inferring Representations of Data in Neural Networks
Finding useful representations of data in order to facilitate scientific knowledge generation is a ubiquitous concept across disciplines. Until the development of machine learning and statistical methods with hidden or latent representations, useful representations of data were generated “by hand” through scientific modeling or simple measurement observations. Scientific models often make explicit the underlying structure of a system which generates the data we observe and measure. To test a model, inferences must be made about the free parameters and the distributions of latent or unmeasured variables in the model conditioned on the data collected. At this time, many scientific disciplines such as astronomy, particle physics, wildlife conservation, and neuroscience have been moving towards collecting datasets that are large and complex enough so that no human will ever look at and analyze all measurements by hand. Datasets of this scale present an interesting scientific opportunity: to be able to derive insight into the structure of natural systems by creating models which can adapt themselves to the latent structure of large amounts of data, often called data-driven hypothesis testing. The three topics of this work fall under this umbrella, but are largely independent research directions. First, we show how deep learning can be used to infer representations of neural data which can be used to find the limits of information content in sparsely sampled neural activity and applied to improving the performance of brain-computer interfaces. Second, we derive a circuit model for a network neurons which implements approximate inference in a probabilistic model given the biological constraint of neuron-local computations. Finally, we provide a theoretical and empirical analysis of a family of methods for learning linear representations which have low coherence (cosine-similarity) and show that linear methods have limited applicability as compared to nonlinear, recurrent models which solve the same problem. Together, these results provide insight into how scientists and the brain can learn useful representations of data in deep and single layer networks.