Blind source separation (BSS) separates a set of desired signals from a mixed observations. Typical applications include sensor array surveillance systems where acoustics sources are to be extracted. However, the nonlinear mixing problem in real environments remains a challenge.
Blind source separation (BSS) separates a set of desired signals from a mixed observations. Typical applications include sensor array surveillance systems where acoustics sources are to be extracted. However, the nonlinear mixing problem in real environments remains a challenge. This project will synergise the use of deep neural networks (DNN) with BSS. The idea is to improve the separation capability of BSS via DNN and reduce the dimensions of the problem for DNN via BSS. As such the performance limitation of BSS linear transformation can be further compensated by DNN through its non-linear reconstruction constraints and BSS will act as a pre-processor, thus allowing DNN a reduced dimension of the problem. This in turn better regulates the numbers of neural nodes and hidden layers required for DNN. An optimization strategy will also be proposed to optimize the DNN via its nodes and layers in response to the parameters in BSS.