Jason Eshraghian

Project Phasor is a community initiative building shared compilers, behavioral virtual machines, and open datasets to scale neuromorphic computing workloads.

Discover how subsampling high-variance pixels in event streams enables 1000x faster visual place recognition for resource-constrained robotics.

Sampling a tiny subset of high-variance event pixels allows robotic visual place recognition systems to successfully operate with low latency and minimal storage.

See how the Neuromorphic Intermediate Representation (NIR) enables seamless SNN model deployment across Norse, SynSense Speck, and SpiNNaker2 hardware.

IBM's NorthPole chip intertwines memory and compute, achieving 25x higher energy and 5x higher space efficiency than conventional architectures.

Discover how the brain could implement gradient descent, addressing weight transport and multiplexing through feedback alignment and node perturbation.

Examining the biological plausibility of gradient descent reveals how node perturbation and feedback alignment successfully bypass the weight transport problem.

The SynSense Xylo hardware and Rockpool toolchain enable the efficient training and direct deployment of spiking neural networks for audio classification.

The Exodus backend accelerates BPTT training of PyTorch-based spiking neural networks in Sinabs before deploying them to the event-driven Speck hardware chip.

See a complete workflow for training spiking neural networks in PyTorch with Sinabs and deploying them directly to the event-driven Speck neuromorphic chip.

The EONS framework applies evolutionary algorithms to co-design spiking neural network topologies and parameters for diverse neuromorphic hardware.

snnTorch applies surrogate gradients to overcome the non-differentiability of spikes, enabling SNN training and quantization using standard PyTorch pipelines.

PEPITA is a biologically plausible learning algorithm that replaces backpropagation with a dual forward-pass mechanism utilizing error-modulated network inputs.