SEQUENCE MODELLING WITH AUTO-ADDRESSING AND RECURRENT MEMORY INTEGRATING NETWORKS
Processing sequential data with long term dependencies and learn complex transitions are two major challenges in many deep learning applications. In this paper, we introduce a novel architecture, the Auto-addressing and Recurrent Memory Integrating Network (ARMIN) to address these issues. The ARMIN explicitly stores previous hidden states and recurrently integrate useful past states into current time-step by an efficient memory addressing mechanism. Compared to existing memory networks, the ARMIN is more light-weight and inference-time efficient. Our network can be trained on small slices of long sequential data, and thus, can boost its training speed. Experiments on various tasks demonstrate the efficiency of the ARMIN architecture. Codes and models will be available.
PDF Abstract
THUMOS14
