How does the Brain recognize speech - Modelling using hierarchical recurrent neural networks

Abstract:

How does the brain recognize speech? In cognitive neuroscience, this question is usually addressed by experiments using neuroimaging methods, e.g. functional magnetic resonance imaging (fMRI) and electroencephalography (EEG). Although there is tremendous progress in better understanding how the human brain recognizes speech, there is actually little progress in elucidating the computational mechanisms of how this is achieved. Here, I present a recently developed computational model which uses recent neurobiological insights from another species, songbirds. Using this computational model, we show that a fusion of two well-established computational approaches, recurrent neural networks and Bayesian filtering, can be used to recognize both birdsong and human speech. The recurrent neural network is based on sequential dynamics as implemented by heteroclinic channels and Hopfield attractor networks. The Bayesian filtering uses a recent formulation which enables online decoding of hierarchical, stochastic, nonlinear dynamical systems. In summary, this model may, on one hand, be an appropriate model for testing quantitative predictions in cognitive neuroscience experiments and, on the other, a novel machine learning tool for artificial speech recognition.


Year: 2012
In session: Hauptvortrag
Pages: 96 to 103