Submit sparse events

Address an event

For neuron n, timestep t, contact k, horizon T, and contacts per neuron K, use flattened indices n*T + t and n*K + k. Both indices must recover the same target neuron by integer division. An event value is a signed floating-point amplitude attached to that exact contact.

Tensor	Shape	Dtype
event_summary_indices	(events,)	int32 or int64
event_contact_indices	(events,)	int32 or int64
event_values	(events,)	Floating point

Keep the tensors in corresponding event order, on the population’s device, and within the declared horizon and contact range.

Execute sparse contact histories

Download population_example.py into the directory where you run this script; build populations explains its construction. The example below uses its three-neuron, four-contact population on CUDA:

import torch
from population_example import build_population

population = build_population("cuda")
summary_indices = torch.tensor([0, 7, 16, 35], device="cuda")
contact_indices = torch.tensor([1, 3, 6, 8], device="cuda")
event_values = torch.tensor(
    [1.0, -0.5, 0.75, 0.25],
    device="cuda",
    requires_grad=True,
)

prediction = population.forward_sparse_contacts(
    summary_indices,
    contact_indices,
    event_values,
    time_steps=12,
)
prediction[:, 4:].square().mean().backward()
print(prediction.shape, event_values.grad.shape)

torch.Size([3, 12, 2]) torch.Size([4])

Replace "cuda" with "cpu" throughout for a CPU run. This example masks the first four causal padding positions before computing the loss.

Gradients and memory

Gradients propagate to event amplitudes, selected contact efficacies, and enabled adaptation banks. Repeated events accumulate in their population-time bins. Dense and sparse paths have an explicit equivalence test in the source repository.

Sparse inputs avoid the dense N*T*K contact-history allocation. Output tensors and the temporal autograd graph still scale with N*T, so increase population size and horizon with their memory cost in mind. Quantized deployment banks have separate interfaces in the synaptic reference and behavior-bank reference.

Next steps

Follow inference-time adaptation to select the banks you want to optimize, or inspect the complete forward_sparse_contacts contract.