Skip to content
§ 04 projects

Projects

Six research projects from the Penn State years plus the current ongoing work at Arm. Each card links to a deeper write-up.

  1. 01 / 07
    ICLR 2025

    NExUME

    Intermittency-aware DNN training for energy-harvesting micro-computers.

    Dynamically adjusts dropout and quantization in response to intermittent energy. DynFit optimizer + DynInfer scheduler. Up to 22% accuracy gain at <5% compute overhead.

    Energy HarvestingIntermittent ComputingDNN TrainingQuantization

    read more →

  2. 02 / 07
    PACT 2025

    Salient Store

    Computational storage for continuous-learning edge servers.

    FPGA-accelerated computational storage devices handle compression, encryption, and redundancy near the data. 6× throughput, 6.1× less data movement.

    Computational StorageContinuous LearningEdgeFPGA

    read more →

  3. 03 / 07
    HPCA 2024

    Usás

    Battery-free continuous learning on solar-powered edge servers.

    A morphable accelerator that adapts to fluctuating solar power, paired with teacher-student training. Hundreds of kWh/year saved per device, no battery required.

    Sustainable ComputingContinuous LearningSolarEdge

    read more →

  4. 04 / 07
    arXiv 2022

    Seeker

    Coreset-based partial inference on energy-harvesting wearables.

    Sensors compute compressed coresets locally and ship only those to the host for final inference. Up to 8.9× communication reduction with near-baseline accuracy.

    EH-WSNHARCoresetEdge

    read more →

  5. 05 / 07
    DATE 2021 ★ Best Paper Nominee

    Origin

    Scheduling and ensemble learning for HAR on energy-harvesting body networks.

    Coordinates intermittent sensors across the body for human activity recognition. 2.5–5% accuracy gain under tight harvested-power constraints.

    EH-WSNHAREnsemble LearningScheduling

    read more →

  6. 06 / 07
    under review

    Prophet

    Neural expert prediction for efficient mixture-of-experts inference.

    Predicts which experts will be activated before the routing layer fires, letting MoE inference avoid round-trips to dormant experts.

    MoELLM InferenceML SystemsEfficient Inference

    read more →

  7. 07 / 07
    ongoing

    SoC Performance & Power Modeling at Arm

    Pre-silicon perf/power analysis for next-generation heterogeneous SoCs.

    Current work on the SoC Architecture Performance team — informing trade-off decisions for compute subsystems targeting Edge AI, IoT, and automotive workloads. Public-facing description; no internal details.

    SoCPerformance ModelingPower ModelingEdge AI

    read more →