Research Projects

Enables HDC on energy-harvesting devices that power-cycle unpredictably. By checkpointing high-dimensional vector state across power interruptions, this work brings ultra-low-energy cognitive computing to batteryless IoT nodes.

Runs DNN inference across multiple parties using Secure Multi-Party Computation so no participant ever sees raw data in plaintext. Protocols are co-optimized with neural architectures to minimize communication overhead while preserving accuracy.

Deploys deep neural networks on microcontrollers with kilobytes of RAM and no GPU. Research spans quantization, pruning, and hardware-aware neural architecture search to maximize accuracy-per-milliwatt on tightly constrained silicon.

Combines energy-harvesting sensors with lightweight vision models to detect crop pests in the field without batteries or connectivity. The system tolerates power interruptions mid-inference, enabling always-available monitoring at near-zero operational cost.

Detects life-threatening ventricular arrhythmias in real time on wearable and implantable cardiac devices under strict microwatt power budgets. Co-designs signal processing and neural classifiers to meet the latency and reliability demands of clinical cardiac monitoring.