A Sparse and Spike-Timing-Based Adaptive Photoencoder for Augmenting Machine Vision for Spiking Neural Networks

Area of Science:

• Neuroscience
• Materials Science
• Computer Engineering

Background:

• Spiking neural networks (SNNs) offer energy-efficient computation by mimicking biological neurons.
• Neural information processing increasingly relies on precise spike timing rather than average firing rates.
• Developing hardware for rapid, sparse, spike-timing-based encoding is crucial for advancing SNNs.

Purpose of the Study:

• To introduce a novel integrated circuit for spike-timing-based encoding of visual information.
• To demonstrate the capability of a 2D material-based photoencoder for SNN acceleration.
• To explore adaptive photoencoding for varying light conditions.

Main Methods:

• Fabrication of a medium-scale integrated circuit using 21 memtransistors based on photosensitive 2D monolayer MoS2.
• Implementation of two cascaded three-stage inverters and one XOR logic gate.
• Encoding of different illumination intensities into sparse spiking signals where time-to-first-spike represents intensity.

Main Results:

• The circuit successfully encoded illumination intensities into spike timing, with higher intensities triggering earlier spikes.
• Demonstrated non-volatile and analog programmability for adaptive photoencoding under scotopic and photopic conditions.
• Achieved low energy expenditure of less than 1 µJ per encoding event.

Conclusions:

• The developed 2D memtransistor-based photoencoder enables efficient spike-timing-based encoding of visual information.
• The bioinspired, in-sensor design accelerates SNNs and offers adaptive capabilities for diverse lighting.
• This technology holds transformative potential for energy-efficient neuromorphic computing.