Neuromorphic Computing Using Long Short-Term Memory Models with Memtransistor-Enabled Short-Term Memory Manipulation

The conventional von Neumann bottleneck, characterized by the separation of information processing and storage, presents architectural limitations in computational efficiency and energy consumption. As the growing demand for energy-efficient computing systems, neuromorphic systems, inspired by the parallel processing principles of the human brain, have gained significant attention. Within this domain, the memtransistor, integrating the functionalities of both a memristor and a transistor into a single device, exhibits distinctive memresistive switching characteristics controlled by an external gate terminal.<br/>To ensure compatibility with CMOS process technology, we fabricated a metal oxide-based memtransistor on SiO₂/p⁺ Si substrate. The memtransistor structure involves the stacking of hafnium oxide (HfO_x) as the gate insulator and indium gallium zinc oxide (IGZO) as the active layer. The redistribution of oxygen vacancies within HfO_x plays a crucial role in controlling the memristive conductance. Introducing an adequate number of oxygen vacancies near the HfO_x/IGZO interface can effectively modulate the Schottky barrier height between the IGZO channel layer and the aluminum (Al) metal source/drain (S/D) electrodes. Unlike traditional two-terminal memristors, the single-pulse measurements reveal the ability to manipulate short-term memory decay time by adjusting gate voltage levels, while the drain pulse waveform remains fixed, allowing us to simulate various degrees of short-term memory effects effectively.<br/>The Long Short-Term Memory (LSTM) model provides distinct advantages for image recognition through controlled degrees of short-term memory states within the LSTM that adapt to varying image dynamics, ultimately enhancing accuracy and efficiency. In time-series data processing, we divide 4x4-sized black-and-white English letter images into four rows, each containing four pixels. The design of electrical pulse waveforms, based on stroke order, serves as inputs to the LSTM mode, enabling it to capture temporal relationships within image sequences. Our research advances memory and recognition technologies, promising enhanced accuracy and efficiency across diverse fields of application.