Title :
A deep architecture with bilinear modeling of hidden representations: Applications to phonetic recognition
Author :
Hutchinson, Brian ; Deng, Li ; Yu, Dong
Author_Institution :
EE Dept., Univ. of Washington, Seattle, WA, USA
Abstract :
We develop and describe a novel deep architecture, the Tensor Deep Stacking Network (T-DSN), where multiple blocks are stacked one on top of another and where a bilinear mapping from hidden representations to the output in each block is used to incorporate higher-order statistics of the input features. A learning algorithm for the T-DSN is presented, in which the main parameter estimation burden is shifted to a convex sub-problem with a closed-form solution. Using an efficient and scalable parallel implementation, we train a T-DSN to discriminate standard three-state monophones in the TIMIT database. The T-DSN outperforms an alternative pretrained Deep Neural Network (DNN) architecture in frame-level classification (both state and phone) and in the cross-entropy measure. For continuous phonetic recognition, T-DSN performs equivalently to a DNN but without the need for a hard-to-scale, sequential fine-tuning step.
Keywords :
convex programming; learning (artificial intelligence); neural nets; parameter estimation; speech recognition; tensors; DNN architecture; T-DSN; bilinear mapping; bilinear modeling; closed-form solution; convex subproblem; deep neural network architecture; hidden representations; learning algorithm; parameter estimation; phonetic recognition; sequential fine-tuning step; tensor deep stacking network; Computer architecture; Error analysis; Neural networks; Speech; Stacking; Tensile stress; Training; deep learning; higher-order statistics; phonetic classification and recognition; stacking model; tensors;
Conference_Titel :
Acoustics, Speech and Signal Processing (ICASSP), 2012 IEEE International Conference on
Conference_Location :
Kyoto
Print_ISBN :
978-1-4673-0045-2
Electronic_ISBN :
1520-6149
DOI :
10.1109/ICASSP.2012.6288994