Long-short term memory (LSTM) is a cognitive architecture that aims to mimic the sequence temporal memory processes in human brain. The state and time-dependent based processing of events is essential to enable contextual processing in several applications such as natural language processing, speech recognition and machine translations. There are many different variants of LSTM and almost all of them are software based. The hardware implementation of LSTM remains as an open problem. In this work, we propose a hardware implementation of LSTM system using memristors. Memristor has proved to mimic behavior of a biological synapse and has promising properties such as smaller size and absence of current leakage among others, making it a suitable element for designing LSTM functions. Sigmoid and hyperbolic tangent functions hardware realization can be performed by using a CMOS-memristor threshold logic circuit. These ideas can be extended for a practical application of implementing sequence learning in real-time sensory processing data.
Bibliographical noteGenerated from Scopus record by KAUST IRTS on 2023-09-23
ASJC Scopus subject areas
- Hardware and Architecture
- Signal Processing
- Surfaces, Coatings and Films