Novel speech processing mechanism derived from auditory neocortical circuit analysis

Analysis of the prominent anatomical and physiological features of auditory thalamus and neocortex has enabled construction of models designed to identify functionality emergent from these biological circuits. These models have recently been shown to provide powerful computational mechanisms for processing of continuous time-varying sequences such as speech; testing on speech databases has yielded positive initial results that are reported here. The model constitutes a novel hypothesis of underlying functions of auditory neocortex, and also represents a novel approach to speech processing. 1. Generalized cortical memory model Research in our laboratory has been concentrating on the phenomenon of long-term potentiation (LTP) [3], which is the most likely candidate for a substrate of neocortical learning and memory. A set of simple learning rules was formulated based on physiological properties of LTP i) synaptic weight can only increase, ii) every increase is small fixed change, and iii) low saturation threshold permits only 5-10 weight increases over the whole period of training [2] [10] . A series of models were constructed based on the known anatomical cortical features sparse-random connectivity in the superficial cortical layers, emergence of the cortical patches defined by the radius of the local inhibition, and feedback inhibition and masking. The above variety of neocortical features specify a biologically constrained class of microcircuits, which typically perform pattern recognition or classification via competitive learning and lateral inhibition [6] [5] . Simulations of those circuits lead to efficient hardware implementations, with a proven utility for pattern recognition via efficient approximation of statistical pattern recognition methods (e.g. Bayes classifiers) [5]. Key anatomical properties of the auditory model being reviewed in [16] (see also Table 1) include i) topographic (MGv) versus broadly-tuned (MGm) thalamic nuclei, convergently projecting to primary auditory cortex; local cortical circuits composed of roughly 100:1 excitatory to inhibitory cells with lateral inhibition; and iii) vertical columnar organization projecting from middle to superficial to deep layers. Key physiological properties include: i) plastic (NMDA-dependent) synapses from broadly tuned MGm afferents versus non-plastic synapsesfrom topographic MGv afferents; ii) plasticity via long-term potentiation (LTP); and iii) time courses for excitation versus inhibition of roughly 1:100. Learning in the model is based on the physiological induction and expression rules for synaptic long-term potentiation or LTP [12] [11] which have been shown in previous modeling efforts to give rise to useful computational properties [2] [10] [9] [5] [7] [8]. Superficial Cortical Layer Broadly tuned afferents from non-specific thalamic nucleus Sparse connectivity ( p ~ .1) LTP of afferent synapses from non-specific nucleus Topographic afferents projecting vertically from the middle layer Topographic vertical projections from superficial to deep layer 100:1 ratio of excitatory to inhibitory cells Middle Cortical Layer Topographic afferents from specific nucleus Equal number of excitatory and inhibitory cells Vertical projections to superficial layer Nonplastic synapses