A Hands-On Course That Turned Undergrads Into Published Researchers

A team of 2nd year CSE students has successfully demonstrated a dynamic ML model-switching approach on smart phones for real-time traffic monitoring. The research paper that resulted has been accepted at a workshop which will be held on the sidelines of the International Conference on Software Architecture (ICSA) 2025. 

When Kriti Gupta, Ananya Halgatti, Priyanshi Gupta, and Larissa Lavanya walked into the Embedded Systems Workshop (ESW) – a hands-on course for 2nd year CSE students that emphasises ‘learning by doing’, they were looking for a novel project to work on. “The previous semester we had an IoT-based course where we worked on a project that allowed us to dabble with sensors and so on. Hence we were looking out for something different,” remarks Ananya. Luckily for them, as part of a collaborative initiative with Qualcomm (Qualcomm EdgeAI lab) to develop EdgeAI use cases at IIITH, Qualcomm has provided developer kits known as the Qualcomm Innovators Development (QID) Kits. These consist of a package of hardware, software, and customer support built on the latest premium SnapDragon system-on-chip (SoC). In the ESW lab itself, there had been a lot of publicity about the kits and the team of 4 grabbed the opportunity to explore and experiment with them. “Not only did we want to do something new but also we wanted to find out how the kits work. Plus, we had heard so much about how the kits can support various transitions,” reasons Kriti. 

What They Did
Prof. Karthik Vaidyanathan of the Software Architecture 4 Sustainability (SA4S) group at the Software Engineering Research Centre has been working on the concept of model switching – that is, when there are a suite of machine learning models deployed on edge devices, a model balancer can intelligently switch between models based on the inputs coming. “As researchers working in the intersection of software architecture and ML, we are constantly trying to improve the efficiency and effectiveness of egdeML systems. One way we attempted to do this is via a self-adaptive mechanism which selects the right model to process data considering operational context and environment such as number of user requests, response time of models, accuracy, energy consumption and so on,” he says. The team of undergrads under the mentorship of PhD student Akhila Matathammal and Prof. Vaidyanathan’s guidance worked on a dynamic model switching approach, which they titled EdgeML Balancer, for object detection on edge devices such as mobile phones. The approach was not only prototyped on the Qualcomm QIDK platform to simulate different scenarios but was also evaluated on real-time traffic data using the smart phone. 

On Edge Now
“Our previous work on model switching was for ML models deployed on the Cloud where we tried to optimise accuracy and latency depending on the situation. But in this case, we experimented on an Edge device which as you know is resource constrained. Here we extended the self-adaptive switching approach to include sustainability in terms of energy consumed by the model,” explains Akhila. With this approach successfully addressing the key challenges currently faced in Edge AI systems, the findings were collated in a paper titled, “EdgeML Balancer: A Self-Adaptive Approach for Dynamic Model Switching on Resource Constrained Devices” and will be presented at the prestigious International Workshop on Flexible Software Architecture for Embedded Computing Systems (SARECS) 2025 in Denmark later this month. 

Research Encounter
Puzzled by all the attention they’re currently receiving, Ananya reveals that they didn’t have a paper publication in mind when they initially embarked upon the project. “It was more of an exploratory project for us and what worked in our favour is that everyone in our team was also interested in learning new stuff,” she says. Calling it a mixed bag, Kriti mentions that the entire experience was equal parts fun and frustration. “The QIDK platform was new and hence we initially had some challenges but with the support of the Qualcomm team and Akhila as well as other seniors like Arya Marda and Shubham Kulkarni who had prior exposure to the QIDK kit, we were able to navigate through them,” she says. 

In The Works
According to Prof. Vaidyanathan, the prototyping phase helped to optimize strategies for Samsung Galaxy M21 smartphone, on which the primary deployment was tested. “We will now be testing it out on the Samsung S24 Ultra and intend to exhibit it at the upcoming R&D Showcase which will be held at IIITH in March,” he says, adding, “The fact that this course work resulted in something substantial beyond mere grades, translating to international recognition is commendable and I hope it serves as inspiration for other UG students too who can see beyond CGPA scores.”