Skip to main content

​
Time is a first-class dimension

In a conventional neural network, a single forward pass processes one sample. In an SNN, a single forward pass processes one sample over T timesteps — time is part of the input shape, not an implementation detail. A sample in THRINDEX has shape [T, n_features]. The output of a model is a spike raster of shape [T, n_output_neurons]. Both dimensions are explicit, fixed at compile time, and stored in the artifact.

​
The timestep dt

dt (delta-t) is the duration of one simulation timestep in milliseconds. It controls how a biological time constant (tau_mem = 20 ms) maps to a discrete recurrence factor: α=e−Δt/τmem\alpha = e^{-\Delta t / \tau_{\text{mem}}} For dt = 1 ms, tau_mem = 20 ms → α ≈ 0.951. For dt = 2 ms → α ≈ 0.905. The same biological time constant produces different discrete dynamics at different dt. dt is declared in the THRINDEX model and stored in the .thx artifact. It is not a runtime parameter. When a model is compiled for a hardware backend with a different native dt, the compiler recomputes alpha and issues a retiming advisory.

​
The spike raster

A spike raster is a binary matrix S[T, N] where S[t, n] = 1 if neuron n fired at timestep t, and 0 otherwise. The raster is the fundamental output of every THRINDEX model. When you run thrindex run model.thx, the printed transcript contains the raster statistics: spike count per neuron, total synaptic operations, prediction (argmax of spike counts).

​
Rate coding vs. temporal coding, revisited

Rate coding reads the raster by column: sum over T for each neuron. The neuron with the most spikes wins the prediction. The timing of individual spikes does not matter.
Temporal coding reads the raster by row: the first timestep at which each neuron fires matters. Neuron 2 firing at t=2 and neuron 5 firing at t=40 encode different values. THRINDEX currently targets rate coding. T = 100 timesteps at dt = 1ms gives a 100ms integration window — sufficient for speech and slow temporal patterns. Models are trained with cross-entropy over spike counts, not over spike times.

​
T and the accuracy floor

Longer T means more integration time and generally higher accuracy — each neuron can accumulate more evidence. But T is a direct multiplier on compute and memory: a model with T=200 does exactly twice the synaptic operations as T=100 for the same input. The practical design rule: use the shortest T that achieves your accuracy target. The keyword-spotting tutorial uses T=100 (100 ms at dt=1 ms) and achieves 64.66% on SHD. Longer T improves this but at linear cost.

​
Synaptic delays

Delays are an additional temporal dimension: a synapse can be configured to transmit its spike with a lag of D timesteps. This allows a network to represent temporal relationships in the input without increasing T. THRINDEX supports per-connection integer delays in the .thx artifact. On hardware backends, delays may be native (the hardware has explicit delay registers) or emulated (the compiler inserts buffer stages). The delay semantics, encoding, and negotiation policy are specified in Concepts: Delays.