Futurists often speak of society’s inevitable technological “singularity,” a point in the near future where computers will become ubiquitous units, seamlessly integrated in everyday objects. This trend is already being foreshadowed by manycore processing via the network-on-chip approach, a novel paradigm which implements on-chip networks that enable platforms with extreme parallel capabilities. Our group seeks to develop new machine learning, optimization, and resource management techniques which can enable such a fundamental shift for energy-efficient, cost-effective, large-scale distributed computational platforms for both embedded and high-performance applications.
Selected Publications
Zero-Shot Neural Architecture Search: Challenges, Solutions, and Opportunities Journal Article
In: IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024.
CANNON: Communication-Aware Sparse Neural Network Optimization Journal Article
In: IEEE Transactions on Emerging Topics in Computing, 2023.
Efficient On-device Training via Gradient Filtering Conference
The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023.
INDENT: Incremental Online Decision Tree Training for Domain-Specific Systems-on-Chip Proceedings Article
In: Proceedings of the 41st IEEE/ACM International Conference on Computer-Aided Design (ICCAD), pp. 1-9, 2022.
Model Elasticity for Hardware Heterogeneity in Federated Learning Systems Proceedings Article
In: Proceedings of the 1st ACM Workshop on Data Privacy and Federated Learning Technologies for Mobile Edge Network (FedEdge), pp. 19-24, 2022.
Domain-Specific Architectures (DSAs): Research Problems and Promising Approaches Journal Article
In: ACM Transactions on Embedded Computing Systems (TECS), 2022.
Embedded Systems
Embedded systems are computer systems that perform dedicated functions while being parts of a larger system. Our research targets primarily heterogeneous many-core System on a Chip (SoC) platforms where communication happens via the network-on-chip. These SoCs should be designed to meet aggressive performance requirements, while coping with limited battery capacity, thermal design power, and real-time constraints. Over the years, we have considered deterministic, probabilistic, and statistical physics-inspired design paradigms. Lately, our research targets machine learning approaches (e.g., imitation and reinforcement learning) for performance and energy optimization and resource management in heterogeneous SoC platforms.
Cyber-Physical Systems
Cyber-physical systems (CPS) refer to a new generation of networked embedded systems that bring together sensing, computation, communication, control and actuation in order to sustain a continuous interaction with the physical world (e.g., processes taking place on electrical power grids, transportation and traffic roads, communication and financial networks, medical devices, smart buildings, etc.). Physical processes are predominantly non-stationary in nature and require time-dependent models for understanding their behavior. Our research focuses on accurate modeling physical processes to better understand the theoretical foundations of CPS design and optimization.