Long Short-term Memory based on a Reward/punishment Strategy for Recurrent Neural Networks

Recurrent neural networks and their variants have received huge success in many difficult tasks, such as handwriting recognition and generation, natural language processing, acoustic modeling of speech, and so on. As a kind of recurrent neural network architectures, the long short-term memory (LSTM) has attracted great attention. Most research works focus on its structures, training algorithms and topology structures. As an improvement to the structure of LSTM, a reward/punishment strategy is developed for LSTM in this paper, which we call RP-LSTM. In RP-LSTM, a reward/punishment (RP) strategy is proposed to evaluate its memory cells' memorization such that it improves its efficiency by forgetting more reasonably. Analysis properties of the developed RP-LSTM is conducted from the neuroscience aspect. To test the performance of the developed RP-LSTM, comparative simulation studies are conducted on three structures, i.e., LSTM, LSTM with forget gate (LSTM-FG) and RP-LSTM. Simulation results on sentiment analysis model and sequence to sequence model demonstrate that RP-LSTM achieves better performance.