| Liquid state machine (LSM) is a computational model that was developed with biological-inspired structure. Its ability to compute lies in the liquid, which is a recurrent neural network that produces non-linear activity. The main idea of the model is that, ideally, LSM is a universal machine given a complex enough liquid. In other words, a complex liquid produces all sorts of non-linearity so that for any given problem, by picking out the useful non-linearity, LSM can solve it.;We take the idea of LSM to a living neural network, with these questions to address. 1) Would a living neural network be a complex liquid that is able to solve non-linear problem? 2) The structure of neural networks in different brain regions are anatomically different and may have different ability to generate non-linear activity that can be utilized for computation. Would we be able to make LSM as a probe to test the computational capability of a given brain region?;Benefitting from high-energy physics technology to detecting neural activity, and the technique of preparing cultured brain slices from different parts of the brain, we are able to study these questions. In my thesis, we fed the living neural network, or the liquid, with electrical current as the input, and train the LSM to compute Boolean functions at a hierarchy of computational difficulty. We show that these non-linear problems can be solved using living neural network to be the liquid, and that the more difficult the task is, the more non-linear activity is required. We also show that in the cortex, the performance of LSM is better in the upper layers than in the deep layers, a possible indication of different computational capability in different brain structures. |
