On-chip learning makes neuromorphic processors adaptable in deployment

Enabling Efficient Processing of Spiking Neural Networks with On-Chip Learning on Commodity Neuromorphic Processors for Edge AI Systems (arxiv 2504.00957) tackles a practical limitation of deployed SNN-based sensing: once trained and deployed, most neuromorphic systems cannot update their weights in the field. Adapting to new target classes or environments requires taking the system offline, retraining, and redeploying.

Why adaptation matters in practice

A counter-UAS seeker faces a continuously evolving target set. Group 1 UAS configurations that were rare two years ago are now common. Modified commercial drones with reduced thermal and acoustic signatures appear regularly. An event-based seeker trained on a fixed target distribution will degrade as the distribution shifts.

The standard solution in machine learning is fine-tuning: collect examples of the new targets, fine-tune the model, redeploy. On an offboard GPU this takes minutes. On an embedded neuromorphic processor with no connection to a training infrastructure, it is currently impractical.

On-chip learning changes this. A processor capable of updating weights from local input-output pairs can adapt continuously in the field without any external infrastructure.

The Loihi 2 implementation

The paper implements on-chip learning on Intel’s Loihi 2 neuromorphic processor — currently the most capable commercially available neuromorphic chip. Loihi 2 has explicit hardware support for spike-timing-dependent plasticity (STDP) and other local learning rules, but using it for practical inference tasks (rather than neuroscience demonstrations) requires careful implementation.

The paper demonstrates an SNN on Loihi 2 that can classify event-camera inputs and update its weights using incoming data during deployment. The adaptation is limited compared to full backpropagation-based training — it can adjust to shifts within the training distribution but not learn entirely new categories — but it is sufficient for the most common adaptation requirement: maintaining performance as target appearance varies within a class.

Power and latency

On Loihi 2, inference plus learning runs at under 500 mW for the target task sizes evaluated. This is within the plausible range for a forward-deployed sensor package, though still above what is achievable with inference-only operation.

The latency impact of learning is minimal: learning happens asynchronously in the background on dedicated cores while inference continues on inference cores. The two workloads are hardware-separated and do not create pipeline stalls.

Limitations and the path forward

On-chip learning on Loihi 2 is constrained to local learning rules — the chip does not implement backpropagation. This limits the classes of network architectures and the amount of adaptation achievable. More capable on-chip learning requires either more capable neuromorphic hardware or hybrid approaches where the local processor does lightweight adaptation and periodic connection to offboard infrastructure handles larger retraining.

For a seeker, the most useful form of this capability may be sensor calibration rather than full model adaptation: periodic recalibration of sensor thresholds and biases based on observed noise statistics, without changing the learned features themselves.

---

Enabling Efficient Processing of Spiking Neural Networks with On-Chip Learning on Commodity Neuromorphic Processors for Edge AI Systems. arxiv 2504.00957