Liquid Computing A Real Effect.pdf

Daniel Goldenholz BME707 Final Project - Boston University Liquid Computing: A Real Effect ABSTRACT A novel approach to temporal pattern recognition was recently introduced in the literature, called liquid computing. The model used was very complex, thus the present study uses a simplified version that is easier to understand. By exploring properties of the system with the simple setup, a better unders tanding of the complex version can be reached. A liquid computer needs a liquid memory and a readout. In the present study, a liquid based on damped coupled oscillators was used. The liquid was a finite element model of water and the wave produced by throwing a rock into it. The readout was a 3 layered back- propagation network. The findings show that even a simplified version of the liquid computer can successfully classify patterns presented serially in time, with or without noise, with or without jitter, and can generalize its training in some cases. It is also shown to correctly learn timing information about its input. The liquid computing framework is therefore shown to be a powerful tool in the analysis of temporal signals using neural networks. BACKGROUND Liquid computing is a technique recently described in the literature (1,2). It provides a theoretical framework for a neural network that can process signals presented temporally. The system is comprised of two subcomponents, the “liquid” and the “readout”. The former acts as a decaying memory, while the latter acts as a the main pattern recognition unit (see Figures 1 and 2). To understand the system, a simple analogy is used. FIGURE 1: The general setup of a liquid state machine. Input comes as a vector into the liquid. The liquid dynamically changes, and is sa