Title: |
Home | Back to Papers |
| Recent experimental data suggest that spike-timing and membrane dynamics of biological neurons may encode information in a way not achievable using artificial neural networks (ANNs) or traditional machine learning algorithms. Practical applications of spike -coded neural networks include flexible and robust artificial intelligence that simultaneously utilize multiple sensory modalities (as do humans) in situation requiring pattern recognition, speech comprehension, and path planning. In particular, continuous speech recognition systems based on hidden Markov models or ANNs have not demonstrated human-like performance levels. To explore the computational capacity and scalability of multimodal neocortex, we have initiated a series of large-scale computer simulations of integrated auditory (spoken sentences) and visual (corresponding lip reading) perception and learning. We distributed more than 25,000 excitatory and inhibitory neurons in layered minicolumns, representing primary sensory and association cortex. Cell compartments included realistic mixtures of voltage-sensitive and calcium-dependent potassium channels. Three types of sentences (modified from the TIMIT corpus) were repetitively presented to the network, and layer-specific Hebbian learning was observed using partial depolarization to serve as a "reward" in association layers integrating auditory and visual spike -coded signals. Preliminary results suggest that computational models of this scope can produce realistic spike encoding o f human speech. | |
| Posted on:31 Oct 2003 | |