Robust signal analysis and recognition techiques for man-machine interaction

点击翻译按钮获取中文摘要

New applications have been developed like automatic speech translation, internet search tools, multi-modal computer interactions. These systems can not adequately process corrupted, noisy signals. They also use pattern recognition techniques that require supervision during training. I propose robust signal analysis techniques and to develop bio-inspired and unsupervised pattern recognizers. Results of the proposed research can be used  to assist in the creation of new intelligent systems that can interact with the environment. Electroneurophysiology can also benefit from the new signal analysis and recognition tools that will be developed for this research. I will work on the development of new systems in sound source separation and intelligent signal recognition with capabilities of being unsupervised and trained on limited data sets. Current state of the art statistical systems usually use supervised training with very much data. Therefore, the design and training of such systems is heavy and very costly. The dynamical, non linear, space and time filtering of spikes combined with neurotransmitters diffusion  along with the topographic organization of neurones yields simultaneous signal processing and recognition. Moreover, spiking allows the encoding of information into a second time scale that is different from the usual time. This second time scale encodes the relative timing of firing neurones. It is a key feature that allows unsupervised training. Synchronization or the generation of specific spiking temporal sequences becomes an auto-organization criteria. On a long term time scale (5 years), I propose to explore the integration and interaction of two fields: the conventional signal and pattern recognition paradigm (discrepancy between analysis and recognition) and the continuum (in organization) and dynamical (in times) analysis/recognition made by the peripheral nervous system with emphasis on signal processing. On a shorter time scale  (2--3 years per project), I will focus on applications like sources separation, signal and speech recognition.