Reproduce “Dendritic Computation through Exploiting Resistive Memory as both Delays and Weights” on OpenResearch

About this paper

Melika Payvand, Simone D'Agostino, Filippo Moro, Yigit Demirag, Giacomo Indiveri, Elisa Vianello

Researchers from the University of Zurich, ETH Zurich, and CEA-LETI developed a neuromorphic hardware architecture that leverages resistive memory (RRAM) to implement both synaptic delays and weights for spatio-temporal processing. This approach achieved up to 95% accuracy in real-time heartbeat anomaly detection while reducing power consumption by up to two orders of magnitude and memory footprint by over 100 times compared to existing recurrent network solutions.

Key insights

• Biologically inspired multi-scale dendritic delays, when physically incorporated into neuromorphic hardware, enable effective spatio-temporal coincidence detection, moving beyond traditional point-neuron models.
• The intrinsic variability of Resistive Random Access Memory (RRAM) in its High Resistance State (HRS) can be leveraged to implement a distribution of passive synaptic delays, transforming a device non-ideality into a functional computational resource.
• Embedding temporal information processing directly into passive synaptic elements allows simpler feed-forward network topologies to handle complex spatio-temporal tasks, reducing the reliance on power-intensive active recurrent connections.

Summary by alphaXiv.