DenRAM: neuromorphic dendritic architecture with RRAM for efficient temporal processing with delays

Area of Science:

• Neuromorphic Engineering
• Computational Neuroscience
• Materials Science

Background:

• Dendritic branching in neocortical pyramidal neurons is crucial for non-linear computation and temporal signal processing.
• Coincidence detection (CD) mechanisms, enabled by synaptic delays, are key for integrating temporally separated inputs.
• Feed-forward spiking neural networks with delays show promise for spatio-temporal pattern recognition, potentially outperforming recurrent architectures.

Purpose of the Study:

• To present DenRAM, the first feed-forward spiking neural network with dendritic compartments implemented in analog circuits using Resistive Random Access Memory (RRAM).
• To demonstrate the capability of RRAM devices to implement synaptic delays and weights for bio-realistic temporal processing.
• To validate DenRAM's efficiency in spatio-temporal pattern recognition and its resilience to hardware noise.

Main Methods:

• Developed DenRAM using analog electronic circuits and RRAM technology on a 130 nm node.
• Configured RRAM devices to emulate bio-realistic synaptic timescales and exploit device heterogeneity for delay implementation.
• Conducted system-level simulations on temporal benchmarks to assess performance and accuracy.

Main Results:

• Experimentally demonstrated DenRAM's ability to replicate synaptic delay profiles and implement CD for spatio-temporal pattern recognition.
• Showcased DenRAM's resilience to analog hardware noise through simulations.
• Achieved superior accuracy compared to recurrent architectures with similar parameter counts.

Conclusions:

• DenRAM introduces advanced temporal processing capabilities to neuromorphic architectures.
• The RRAM-based design offers a reduced memory footprint for edge devices and high accuracy on temporal tasks.
• DenRAM represents a significant advancement in low-power, real-time signal processing technologies.