Dimensionality reduction to maximize prediction generalization capability

Generalization of time series prediction remains an important open issue in machine learning, wherein earlier methods have either large generalization error or local minima. We develop analytically solvable, unsupervised learning scheme that extracts the most informative components for predicting future inputs, termed predictive principal component analysis (PredPCA). Our can effectively remove unpredictable noise and minimize test through convex optimization. Mathematical analyses demonstrate that, provided with sufficient training samples sufficiently high-dimensional observations, PredPCA asymptotically identify hidden states, system parameters, dimensionalities canonical nonlinear generative processes, a global convergence guarantee. performance using sequential visual inputs comprising hand-digits, rotating 3D objects, natural scenes. It reliably estimates distinct states predicts outcomes previously unseen input data, based exclusively on noisy observations. The simple architecture low computational cost are highly desirable neuromorphic hardware.