Dynamic Firing on Static Analog/Digital Neuron Circuits with Resistive Synapses for Time-Series Neural Network

• Main Authors: Marukame, Takao, Sugino, Junichi, Kitamura, Toshimitsu, Ishikawa, Kazuo, Takahashi, Koji, Tamura, Yutaka, Nomura, Kumiko, Mizushima, Koichi, Nishi, Yoshifumi
• Format: Conference Proceeding
• Language: English
• Subjects: Analog-Digital Neuron Circuits; Artificial neural networks; Delays; Firing; Memristive Devices; Neurons; Nonlinearity; Reservoir Computing; Resistors; Semiconductor device measurement; Spike-Timing-Dependent Plasticity (STDP); Synapses; Time-Series Neural Networks
• ISSN: 2158-1525; 2158-1525
• DOI: 10.1109/ISCAS45731.2020.9180642

An analog-to-digital mixed circuit with resistors for static neurons was implemented on a CMOS IC chip. By comparing current magnitudes via the resistors on two lines, the neuron circuit output a multiplier accumulation result and a step function at the comparator as firing. Analog neurons with 1024 synapses and digital peripherals always operated at around 10 mW. The firing delay was intrinsically caused by patterns of inputs and synaptic weights. A recurrent connection directly from the output to the input generated an oscillation, and thus average latency of about 1 μs could be estimated from the observed period. Dynamic firing was observed even with digitally controlled recurrence, indicating a data-converter function from static to dynamic in which the firing pattern can be tuned via randomized synapses. A potential application is reservoir computing, where nonvolatile memristive devices can be further added for readout and learning with timing-dependent plasticity.