Modern embedded AI systems rely on Deep Neural Networks (DNNs) running on resource-constrained devices such as wearables, smart sensors, hearables, and IoT nodes. While current deployment methodologies can optimize models before deployment, the resulting software remains static throughout the lifetime of the device. Consequently, energy consumption, memory usage, and computational requirements remain fixed, even when the application or environmental conditions evolve over time.
This project investigates a radically new paradigm for Embedded AI: enabling devices to dynamically compress and adapt neural networks directly on-device after deployment. Inspired by how humans continuously optimize their physical efficiency through training, this project aims to develop software methods that allow deployed neural networks to progressively reduce their resource footprint while maintaining application performance.
The research focuses include: runtime-reconfigurable neural network software kernels for novel embedded systems architecture, performance models for adaptive AI workloads, on-device search and optimization engines for memory-aware deployment, compiler and runtime systems capable of supporting dynamic neural network topologies, energy-efficient embedded AI frameworks targeting RISC-V multicore and heterogeneous computing platforms.
The project combines fundamental research in embedded software and compiler systems with practical validation on state-of-the-art embedded AI hardware platforms. Successful outcomes could significantly extend battery lifetime and enable a new generation of adaptive and sustainable embedded AI systems
