Hidden Markov models (HMMs) have been extensively used in the engineering community for temporal pattern recognition (see Thrust area II(B)). Perhaps the most sophisticated use of such models has been in speech recognition systems, and specifically in ``word-spotting'' applications where the HMMs are decoded in a multi-pass algorithm: state likelihoods are computed for an entire spoken phrase, followed by a ``backtracking'' pass to determine the spoken words. While this multi-pass approach is efficient computationally, it is highly non-biological. Instead, an interesting alternative architecture uses hidden Markov models in an on-line, single-pass fashion [.Lippmann 1994.], where state likelihoods are computed for each frame, and recognized temporal sequences are flagged based only on these current-frame likelihood values.
It is possible to map this online HMM algorithm to continuous-time analog VLSI hardware [.Lazzaro 1997.]. Here we propose to investigate the mapping of the algorithm to biologically motivated cortical circuitry similar to those developed in detail in [.Byrne Shamma 1997.]. By understanding plausible strategies for computing HMM decoders in neural hardware, we hope to better understand how biological auditory processing systems combine front-end feature detection computation with later temporal processing computations [.Lazzaro 1997, Lazzaro 1994.]. These HMM algorithms and implementations are also extremely relevant for their effective transitioning into actual applications in manufacture process control and machine fault monitoring ( Thrust area V).