ABSTRACT
Temporal information processing is critical for a wide spectrum of applications, such as finance, biomedicine, and engineering. Reservoir computing (RC) can efficiently process temporal information with low training costs. Various memristors have been explored to demonstrate RC systems leveraging the short-term memory and nonlinear dynamic behaviours. However, the short-term memory is fixed after the device fabrication, limiting the applications to diverse temporal analysis tasks. In this work, we propose the approaches to modulating the short-term memory of Pt/SiOx:Ag/Pt memristor for the performance improvement of the RC systems. By controlling the read voltage, pulse amplitude and pulse width applied to the devices, the obtainable range of the characteristic time reaches three orders of magnitude from microseconds to around milliseconds. Based on the fabricated memristor, the classification of 4-bit pulse streams is demonstrated. Memristor-based RC systems with adjustable short-term memory are constructed for time-series prediction and pattern recognition tasks with different requirements for the characteristic times. The simulation results show that low normalized root mean square error of 0.003 (0.27) in Hénon map (Mackey-Glass time series) and excellent classification accuracy of 99.6% (91.7%) in spoken-digit recognition (MNIST image recognition) are achieved, which outperforms most memristor-based RC systems recently reported. Furthermore, the RC networks with diverse short-term memories are constructed to address more complicated tasks with low prediction errors. This work proves the high controllability of memristor-based RC systems to handle multiple temporal processing tasks.
ABSTRACT
Reservoir computing is a powerful neural network-based computing paradigm for spatiotemporal signal processing. Recently, physical reservoirs have been explored based on various electronic devices with outstanding efficiency. However, the inflexible temporal dynamics of these reservoirs have posed fundamental restrictions in processing spatiotemporal signals with various timescales. Here, we fabricated thin-film transistors with controllable temporal dynamics, which can be easily tuned with electrical operation signals and showed excellent cycle-to-cycle uniformity. Based on this, we constructed a temporal adaptive reservoir capable of extracting temporal information of multiple timescales, thereby achieving improved accuracy in the human-activity-recognition task. Moreover, by leveraging the former computing output to modify the hyperparameters, we constructed a closed-loop architecture that equips the reservoir computing system with temporal self-adaptability according to the current input. The adaptability is demonstrated by accurate real-time recognition of objects moving at diverse speed levels. This work provides an approach for reservoir computing systems to achieve real-time processing of spatiotemporal signals with compound temporal characteristics.
ABSTRACT
Memory-augmented neural network (MANN) has received increasing attention as a promising approach to achieve lifelong on-device learning, of which implementation of the explicit memory is vital. Content addressable memory (CAM) has been designed to accelerate the explicit memory by harnessing the in-memory-computing capability. In this work, a CAM cell with quadratic code is proposed, and a 1Mb Flash-based multi-bit CAM chip capable of computing Euclidean (L2) distance is fabricated. Compared with ternary CAM, the latency and energy are significantly reduced by 5.3- and 46.6-fold, respectively, for the MANN on Omniglot dataset. Besides, the recognition accuracy has slight degradation (<1%) even after baking for 105 s at 200°C, demonstrating the robustness to environmental disturbance. Performance evaluation indicates a reduction of 471-fold in latency and 1267-fold in energy compared with GPU for search operation. The proposed robust and energy-efficient CAM provides a promising solution to implement lifelong on-device machine intelligence.