ICANN · 2026 · First author accepted

CLANE: continual learning of actions from event cameras on Loihi 2

An end-to-end on-chip pipeline that learns new human actions from an event-camera stream, class-incrementally, without sending data off the device.

16×lower latency than a Jetson Orin Nano pipeline
>100×lower energy than the same baseline
5 msper sample, at mJ-level energy

CLANE pipeline: a spiking CNN is pretrained on basic actions on a GPU, then frozen on Loihi 2 where temporal aggregation, normalization and a CLP-SNN head learn novel actions online from the event-camera stream. — CLANE: a spiking CNN pretrained offline, then class-incremental learning of new actions on Loihi 2 with a CLP-SNN head. Figure from the paper.

Problem

Recognizing and learning new human actions on the device, from a sparse, asynchronous event-camera stream, without sending data off the device. On-device learning matters here for both privacy and low-latency adaptation.

Approach

An end-to-end on-chip pipeline: a 2D spiking CNN feature extractor, temporal aggregation, normalization and a CLP-SNN head for class-incremental learning. The Temporal Aggregation and fixed-point Normalization layers are new Loihi 2 modules built for this work.

Evaluation

THU-EACT-50 (50 classes), with iso-algorithm cross-platform benchmarking at three evaluation levels against a sequential CNN + GRU + CLP pipeline on a Jetson Orin Nano.

Result

70.4% accuracy in the continual-learning task at 5 ms per sample and mJ-level energy: 16× lower latency and more than 100× lower energy than the Jetson Orin Nano baseline. With a co-author, I also built a live demo of real-time action learning on Loihi 2 from a physical DVS event camera.

Paper

CLANE: Continual Learning of Actions on Neuromorphic Hardware from Event Cameras

Elvin Hajizada, Michael Neumeier, Edward Paxon Frady, Yulia Sandamirskaya, Axel von Arnim, Bing Li, Eyke Hüllermeier · † equal contribution

ICANN 2026 · accepted