Detailed  Schedule

Time: 08:00 AM – 09:00 AM  

Registration

Time: 9:00 AM – 10:00 AM  

Technical Session I (Invited Talk): 

Speaker: Prof. Nirmalya Roy  (University of Maryland, Baltimore County )

Speaker Website: https://mpsc.umbc.edu/

Lecture Title: Designing Camera-based Physiological Health Parameter Sensing on the Edge

Session Chair: Dr. Jaydeep Howladar, NIT Durgapur

Title of the Talk: Designing Camera-based Physiological Health Parameter Sensing on the Edge

Abstract: Contactless physiological health parameter sensing using video cameras provides a low-cost, ubiquitous, and easy-to-deploy solution for continuous health monitoring and assessment. Remote photoplethysmography (rPPG) and respiration rate estimation, for instance, enable measurement of two key human vital signs: heart rate (HR) and respiratory rate (RR). These vital signs can be extracted from facial and respiration-induced chest movement videos using regular camera sensors. However, traditional signal processing methods for deriving these vital signs from videos often rely on heuristics and fail to harness the power of data-driven deep learning (DL) models. In this talk, I will articulate the challenges in developing robust data-driven models for rPPG estimation, particularly the inherent aleatoric (irreducible) uncertainty in ground truth data annotation. This uncertainty arises due to factors such as synchronization issues between source (camera) and target (ground truth) data streams, sensor placement and heterogeneities, and inter-subject variations. These challenges hinder the development of traditional end-to-end DL models, potentially leading to overfitting on irrelevant noise in the absence of clean labels. To address these uncertainties and build robust DL models for rPPG estimation, I will discuss three novel approaches: (a) Multi-task Learning: A method that separates individual targets while learning shared embeddings across noisy ground truth data to identify core rPPG features, (b) Self-Supervised Learning: An approach that leverages unlabeled rPPG data to learn intrinsic rPPG properties and enhance feature representation and signal reconstruction by incorporating prior domain knowledge (e.g., HR frequency, phase, and temporal coherence), (c) Generative Adversarial Learning: A framework that enables fine-grained rPPG feature learning without direct supervision, leveraging large-scale rPPG datasets. I will also present validations of these approaches on our in-house MPSC-rPPG dataset as well as multiple public datasets, with open-source code for further exploration. Additionally, I will demonstrate RhythmEdge, a real-time heart rate estimation system built using a low-cost camera sensor and discuss rPPG-specific pruning techniques to optimize model size for efficient edge deployment. The talk will conclude with future research directions and insights into our ongoing work on respiratory rate estimation using a particle-video-based approach.

 Title of the Talk: Efficient Machine Learning for the Edge: Resolving Challenges of Labeling, Collecting, and Collaborating Information Across Devices 

Efficiency is one of the key design aspects for bringing machine learning (ML) to the edge, where resource constraints and scalability challenges prevail. This talk explores efficiency from three key perspectives: labeling efficiency, minimizing data collection efforts, and fostering collaborative learning across devices. First, I will present an innovative approach leveraging state-of-the-art large language models (LLMs) to automate annotation tasks in sensor data, significantly reducing the human effort that is usually required to label any dataset. Next, I will try to introduce an idea for uncovering hidden classes within audio datasets, demonstrating how to make the most out of existing datasets, thus reducing the efforts to perform new data collections for every new application. Finally, I will introduce a robust framework for deploying federated learning across resource-constrained devices, enabling collaborative model training across multiple edge devices while preserving privacy and reducing communication overhead. Together, these contributions highlight the importance of interdisciplinary methodologies in driving efficiency and scalability in machine learning systems, paving the way for broader accessibility and practical deployment. 

Time: 12:30 PM – 1:00 PM :  Sponsored Key Note

1:00 PM-2:00 PM : LUNCH

Time: 2:00 PM – 3:30 PM : Joint Keynotes for all workshops (Dilip Krishnaswamy, Kiran Kuchi) 

Time :3:30 PM – 4:30 PM:  Paper Presentation (3 Papers)

Session Chair: Dr. Suoy Saha, NIT Durgapur

4. Non-Fungible Tokens (NFTs): A Systematic Study of Trust Criteria

Rangin Lahiri (UEM Kolkata),  Saikat Chakrabarti (UEM Kolkata) and Subrata Saha (UEM Kolkata)

5. Faster Convergence of Adaptive Data Rate in Dynamic LoRa-based IoT Network Applications

Alekhya Gorrela (BITS Pilanai Hyderabad) and Nikumani Choudhury (BITS Pilanai Hyderabad)

6. Optimizing Autonomous Intersection Control Using Single Agent Reinforcement Learning

Yash Ganar (RGIPT Amethi), Vinay Kumar (RGIPT Amethi), Shraddha Dulera (IITRAM Gujrat) and Ram Narayan Yadav (IITRAM Gujrat)

Time: 4:30 PM : Best Paper Award