Fully asynchronous, event-driven neural network for neuromorphic processor

Fully asynchronous, event-driven (AED) neural networks in JAX + MPI.

Each MPI rank owns one layer (or a shard of one layer). Layers communicate by passing sparse events — (neuron_idx, value) for MLP, (channel, x, y, value) for CNN/ResNet — through MPI point-to-point messages, terminated by an END_SIGNAL. Neurons fire and forward individually rather than waiting for a whole layer to finish.

Runners

Script	Model
async_MLP_general.py	AED multilayer perceptron
async_CNN_general.py	AED convolutional network (data + model parallelism)
async_ResNet_general.py	AED CNN with residual (skip) connections

Running

mpirun -n <ntasks> python async_MLP_general.py --config configs/MLP_config.yaml

<ntasks> must be a multiple of len(layer_sizes); ntasks / len(layer_sizes) is the number of data-parallel replicas. For inference from a checkpoint, set mode: inference and rerun: <path/to/checkpoint.json> in the config.

One generic config per runner lives in configs/ — every supported parameter is listed there with a comment.

Key hyperparameters

• firing_nb — top-k neurons that fire per layer per event. Lower means fewer events in the network: faster and more energy-efficient, at some accuracy cost.
• sync_rate — how many events a neuron must receive before it may fire again. Use 1 when benchmarking raw performance.
• restrict — soft-reset multiplier applied to a neuron's value after it fires.
• frame_size — event datasets only: fire the first hidden layer once per true time frame instead of using its sync_rate.

Layout

async_{MLP,CNN,ResNet}_general.py   runners
configs/                            one generic example config per runner
dataset_helpers/                    dataset loaders (MNIST, N-MNIST, SHD, DVS, NCARS, CIFAR-10, iris)
forward_backward_pass/              event-driven inference, backprop, losses
other_helpers/                      MPI partitioning, params/config handling, weight init, pooling

Results are written to network_results/<dataset>/training/<arch>/ as JSON. Note that result JSON keys are space-separated (firing number, synchronization rate, learning rate), not the snake_case config names.