Speaker:Jie Yuan

Location: WebEx

Sequence Prediction and Classification Based on Long- and Short-Term Memory Retrieval Architectures

Applying traditional recurrent neural networks (RNNs) on extremely long sequential data is infeasible due to the high time complexity and the limited capability of the memory units in RNNs. To alleviate this problem, we propose a new deep neural network-based architecture named Long- and Short-term Memory Retrieval (LSMR) architecture for sequence prediction and classification. LSMR architecture consists of a query extractor, a long-term memory retriever, and a predictor. The query extractor and the long-term memory retriever compose a long-term memory retrieval mechanism that enables the LSMR to handle extremely long sequences. Experiments on our WSJ news corpus and human activity recognition dataset demonstrate the superior performance of our proposed models, compared with long-term or short-term memory-based models.

Committee: Zhu Zhang (major professor), Qi Li, Sheng Bao, Jin Tian, and Wallapak Tavanapong