Cristian Borcea

Professor

borcea@njit.edu

Research Areas: Mobile systems, federated learning

Privacy-Preserving Federated Multi-Task Learning

Multi-task learning (MTL) enables simultaneous learning of related tasks, enhancing the generalization performance of each task and facilitating faster training and inference on resource-constrained devices. Federated Learning (FL) can further enhance performance by enabling collaboration among devices, effectively leveraging distributed data to improve model performance, while ensuring that the raw data remains on the respective devices. However, conventional FL is inadequate for handling MTL models trained on different sets of tasks. We propose FedMTL, a privacy-preserving FL aggregation technique that handles task heterogeneity across users. FedMTL generates personalized MTL models based on task similarities, which are determined by analyzing the parameters for the task-specific layers of the trained models. To prevent privacy leakage through these model parameters and to protect the privacy of the task types, FedMTL employs low-overhead secure multi-party computation aggregation.

Kasthuri Jayarajah

Assistant Professor

kasthuri.jayarajah@njit.edu

Research Areas: Sensing, Robotics, Context Awarenesse, Mobile Systems, Edge AI

Human-Robot Cooperative Systems

Effective communication between humans and robotic teammates is imperative for successful cooperation. However, direct communication (e.g., verbally) can be distracting to humans under high cognitive workload scenarios. Thus, we explore the feasibility of using multimodal wearable sensing to facilitate implicit coordination where robots are able to synchronize with their human teammates without explicit intervention. We have developed systems that extract human intent and task performance using physiological sensing, visual attention and gestures for cooperation in scenarios such as search and rescue, teleoperation and cognitive rehabilitation, addressing key system challenges in dealing with varying sensing fidelity, hardware resource constraints and stringent latency bounds

Collaborative Edge Intelligence

While collaborative intelligence over a network of sensors can lead to more accurate inferencing as opposed to isolated, individual inferences, the cost of collaboration (e.g., increased bandwidth requirement) can be highly prohibitive. The problem is especially exacerbated for richer data representations such as images and point clouds. We developed various architectures for collaboration for multi-view camera systems and heterogeneous multi-robot systems investigating the performance trade-offs of early and late fusion methods and more recently, Com AI, where we adapt deep learning pipelines on-the-fly based on lightweight digest sharing between collaborating nodes. ComAI incurs zero re-training of the deep networks, adds minimal overhead to the processing pipeline as well as the bandwidth required for sharing digests over the network. We further explore selfconfiguration and steerability for collaborative systems using Reinforcement Learning (RL) where cameras learn to identify regions that benefit the most from collaboration, quantify the goodness of collaborators for selective collaboration and over time, “steer” themselves towards optimal configurations.