Locally Competitive Algorithms for Sparse Approximation

摘要: Practical sparse approximation algorithms (particularly greedy algorithms) suffer two significant drawbacks: they are difficult to implement in hardware, and inefficient for time-varying stimuli (e.g., video) because produce erratic temporal coefficient sequences. We present a class of locally competitive (LCAs) that correspond collection principles minimizing weighted combination reconstruction MSE cost function. These systems use thresholding functions induce local nonlinear competitions dynamical system. Simple analog hardware can the required nonlinearities competitions. show our LCAs stable under normal operating conditions sparsity levels comparable existing methods. Additionally, these coefficients video sequences more regular (i.e., smoother predictable) than produced by algorithms.