講演情報
[8p-S102-5]Physical Reservoir Computing Using Artificial Synaptic Devices
〇Hongseok Oh1 (1.Soongsil Univ.)
キーワード：
Physical Reservoir Computing、Artificial Synaptic Devices
Physical reservoir computing (PRC) is a new computing framework, which is developed for energy-efficient machine learning of time-series data. Usually, a recurrent neural network (RNN) is used for learning time-series data, which is powerful but requires high computational cost. Reservoir computing (RC) is developed to address the intensive computational cost. In RC, a reservoir, composed of fixed, randomly connected nodes, is driven by the input signal. Only the connections between so-called readout nodes and output nodes are trained, while most of the connections remain unchanged. By minimizing the number of connections for training, this approach can significantly reduce computational costs. On the other hand, the performance is maintained comparable or even superior compared to RNN.
PRC is an approach where the digital reservoir is replaced with a physical system. Here, the energy efficient AI can be implented since reservoir requires no computational cost. Any physical systems with nonlinearity and fading memory characteristics can serve as a PRC reservoir, including mechanical, electrical, or optical systems. Recently, PRC systems based on one or a few (opto)electronic devices have emerged as practical solutions, thanks to their simplicity and compatibility with existing electronics.
In this presentation, we introduce tellurium thin-film-based photonic synapses and their application to PRC. We demonstrate two important examples: a classification task of grayscale handwritten digits, and prediction task of nonlinear dynamical equations. Our research can enable energy-efficient AI for learning and processing dynamic systems in the future.
PRC is an approach where the digital reservoir is replaced with a physical system. Here, the energy efficient AI can be implented since reservoir requires no computational cost. Any physical systems with nonlinearity and fading memory characteristics can serve as a PRC reservoir, including mechanical, electrical, or optical systems. Recently, PRC systems based on one or a few (opto)electronic devices have emerged as practical solutions, thanks to their simplicity and compatibility with existing electronics.
In this presentation, we introduce tellurium thin-film-based photonic synapses and their application to PRC. We demonstrate two important examples: a classification task of grayscale handwritten digits, and prediction task of nonlinear dynamical equations. Our research can enable energy-efficient AI for learning and processing dynamic systems in the future.