Hypnos divides continual learning between a RISC-V edge device and a smart charger: the device makes lightweight, gradient-free updates and retains compressed representations, while the charger uses replay and gradient updates to consolidate the model. The researchers report benchmark results for the system, including accuracy and device-side latency figures that depend on activation precision.
How Hypnos splits learning between device and charger
Continual learning allows a model to adapt as it encounters new data. In Hypnos, the edge device does not update the entire model with error backpropagation. It keeps a frozen MobileNetV3-Large feature extractor, updates an online prototypical classifier, and encodes representations using product quantization, a way to store them in compressed form. The design uses a RISC-V microcontroller.
The smart charger takes on the more resource-intensive phase. It decodes the retained representations, replays them during gradient-based updates, and returns consolidated parameters to the device. The described architecture does not require a cloud training server.
Device-side adaptation and charger-side consolidation are separate stages, so a device-latency figure describes only one part of the learning cycle.
What the reported evaluation and results measure
The paper abstract reports that Hypnos was deployed on a ZCU104 FPGA and synthesized in a 28-nanometer process. Those are the stated research evaluation conditions. The abstract reports results on the CORe50 NICv2-391 benchmark: accuracy ranges from 78.6% to 81.6%, and on-device latency from 7.79 to 10.54 seconds per sample, depending on activation precision. It does not pair every endpoint in either range with a specific precision. The Hypnos paper abstract
For the INT8-activation configuration, the reported result is 78.6% accuracy and 7.79 seconds per sample of device-side latency. That latency excludes consolidation on the charger, so it is not a timing for the complete two-stage cycle.
Against quantized latent replay, the researchers report 3.8 times fewer learning steps and 22 times less learning compute, alongside higher accuracy and comparable latency. The implementation report also gives 33.3 mW for continuous-operation power in synthesis and a 5.48 MB replay buffer on the charger. The power figure is a synthesis result; the buffer is allocated to the charger.