Data compression by recurrent neural network dynamics
Leong Kwan Li · 2002
Data compression is dominated by the Fourier or wavelet transforms which approximate the given function or sequence as a linear sum of the basis functions. In this paper, we discuss the use of dynamical systems for compression. Since leaky-integrator model neural nets can approximate arbitrary finite sequences, we propose to compress a 'not too wild' signal by a recurrent neural network. As an initial valued problem, the information to be stored are the parameters of the system and the initial states. Elementary analysis on error and compression ratio are also given.