Peking University School of Integrated Circuits has made important progress in the research of deep pool computing hardware

Recently, the team led by Academician Huang Ru and Professor Yang Yuchao from the School of Integrated Circuits at Peking University has made important progress in the research of deep reservoir computing hardware.

The experimental research results show that compared with single-layer reservoir computing systems of the same size, the deep reservoir computing hardware constructed in this work has larger memory capacity, richer number of reservoir states, and hierarchical information processing capabilities. It also demonstrates excellent performance in tasks such as waveform classification and power consumption prediction, proving the potential of deep reservoir computing systems in time series information processing tasks.

According to the news from Peking University Microelectronics, the relevant results were published in Advanced Materials under the title of "Multilayer Reservoir Computing Based on Ferroelectric α-In2Se3 for Hierarchical Information Processing". Liu Keqin, a 2018 doctoral student at the School of Integrated Circuits of Peking University, is the first author, and Professor Yang Yuchao and Academician Huang Ru are the corresponding authors.

Reservoir computing is a recurrent neural network (RNN) with low training cost and low hardware overhead, which has a wide range of applications in time series information processing, such as waveform classification, speech recognition, time series prediction, etc. At present, international research on reservoir computing systems mainly focuses on exploring the use of different types of nonlinear devices (such as memristors, spin torque oscillators, nanowire networks, semiconductor optical amplifiers, etc.) to build a single-layer reservoir, but the limitations of the number of reservoir states, memory capacity, and complex dynamic characteristics fundamentally restrict the improvement of the system's own information processing capabilities.

In response to this key issue, the research team led by Academician Huang Ru and Professor Yang Yuchao of the School of Integrated Circuits of Peking University has for the first time used cascaded short-term nonlinear units to build deep reservoir computing hardware. By increasing the number of reservoir layers, it has achieved hierarchical information processing capabilities, richer reservoir states, larger memory capacity, and more complex dynamic characteristics. This work has taken an important step in building a deep reservoir computing hardware system and realizing hierarchical temporal information processing.