Is the Liquid State Machine Model of a Cortical Minicolumn Compatible with a Tonotopic Map?
Liquid state machines (LSMs) have recently been proposed as a model of the cortical minicolumn in the brain. They consist of random, recurrent circuits of excitatory and inhibitory neurons whose temporal dynamics resemble a viscous liquid. Such networks provide good support for nontrivial pattern recognition tasks. It is known that cortical columns are also arranged in topographic maps, such as retinotopic for the visual modality, and tonotopic for the auditory modality. The present research explores whether the LSM architecture is compatible with the use of topographic maps in the auditory modality. We studied pattern recognition performance on a spoken digit recognition task, comparing LSM architectures that were, and were not, embedded in a tonotopic map. We found that LSM performance was degraded when embedded in a tonotopic map. We conclude that the LSM model is deficient when embedded in a broader context of cortical function.
