Reservoir computing Overview

Recently, I am mainly looking at reservoir computing. The RC field is jointly established by two twin paper articles. The two paper documents propose the echo state network concept from the engineering perspective, another concept of liquid state machine is proposed from bio's perspective. The links to the two paper documents are as follows. If you are interested, refer to them:

The "Echo state" approach to analyzing and Training recurrent neural networks

Real-time computing without stable states: A New Framework for neural computation based on Perturbations

 

I have also seen a good overview recently, an overview of reservoir computing: Theory, applications and implementations.

The concept of reservoir computing is simply to give some input to a randomly connected spiking neural network (called reservoir, then we use the existing machine learning method to learn the pattern of the network for the specific input, so as to achieve the pattern recognition goal. We can make an analogy. If we equate SNN with the circuit of the brain, then reservoir computing is to extract several outputs from the brain circuit, then, determine the pattern activities of these neurons connected to the output and classify them to determine the information received by the input (this is a bit like the brain computer interface method using EEG ). Of course, the reservoir model can be not limited to SNN, but also the node model can be varied, including the polychronization of Izhikevich mentioned earlier, some people use it as reservoir to do some Pattern Recognition Research (for details, see delay learning and polychronization for reservoir computing ).

I personally think that RC is another field that can produce a large number of applications since the traditional ANN, and it also reveals a form of neural tissue in the brain to a certain extent. However, if you only focus on the research of its application, it can be said that it is a way to stop. A small number of independent neuron groups cannot simulate advanced brain functions. The structure of the randomly dispersed neuron may be correct in a small local area, but it is not completely consistent with the feed-forward feedback between the hierarchical processing structure and structure in the brain. Of course, RC can be a good pitfall, and research around it may reveal more about the principles of the brain.

Supplement: I recently saw several paper articles by Stephen Grossberg from Brown University. I have analyzed how visual cortex neuron structure produces stereoscopic vision, visual illusion, and other phenomena. I have read this article, I plan to study it in the next stage.

The following are links and tools related to RC that have been viewed over this period of time. You can organize them for your reference:

1. http://www.hirnforschung.net/cneuro/ is a site on Computational neurology and neuroinformatics summary, which includes a large number of neural simulation tools

2. csim is a brain Circuit Simulator Based on C ++ using Matlab Mex interface. This website is developed by Henry, the leader of the blue Brain Project, and Wolfgang, who proposed the liquid state machine concept.

3. http://reslab.elis.ugent.be/software is two reservoir computing toolbox, one is written in MATLAB, the other is written in Python, can come to play when you are free.

 

Continue to be the header of MATLAB Mex and python...

 

 

Keywords: reservoir computing, brain-like Algorithm