E0401: LIF layer cannot be mapped to akida-akd1500
What happened
A .thx artifact was validated against the akida-akd1500 backend and one of its layers has type: "lif". The AKD1500 hardware does not implement leaky integrate-and-fire (LIF) dynamics and cannot execute this layer.
Why
AKD1500 implements Akida 1.0. Its activation function is a bounded ReLU applied per-inference-call to an integer dot product. It has:
- No membrane potential — there is no state variable that accumulates across time.
- No exponential leak — the
alpha = exp(-dt / tau_mem) term in the LIF equation has no hardware equivalent.
- No spike-triggered reset — the subtract or zero reset modes have no hardware equivalent.
LIF neurons and bounded ReLU are different computational models. There is no lossless conversion between them.
How to fix
Option A — Use the simulator. If you need SNN inference with LIF dynamics, compile for target="sim":
The simulator implements LIF fully, including membrane potential, synaptic dynamics, and both reset modes.
Option B — Redesign for AKD1500. AKD1500 runs feedforward integer networks (Dense → ReLU). If you want hardware acceleration and do not need SNN temporal dynamics, redesign the model without LIF layers. Refer to the BrainChip documentation for the supported layer types.
There is no automatic model conversion. A model that uses LIF neurons depends on temporal state that does not exist in the AKD1500 execution model.
Example
