Tags:
Muhammad Zawish, PhD Research Student, describes what it is like working as part of Walton’s team.
How long have you been working in Walton Institute
I joined Walton Institute as a funded PhD Research student in March, 2020.
Outline your day to day
Being a second year PhD student, one of my main tasks is to read and analyse technical papers to find potential gaps in the domain. I also code and develop AI models for resource-constrained execution, particularly for the application of Smart Farming. Usually I spent the morning reading literature, while I focus on development tasks during afternoon.
How and why you got into your research field
During my undergrad in Computer Engineering, I got involved in some small research-based projects and was also lucky to be able to publish them. Since then, I decided to pursue a research-oriented career in computer science, particularly in AI. For this reason, I got an opportunity at Walton as a PhD Student to develop lightweight AI algorithms for smart agri applications under the VistaMilk project.
Briefly outline the research projects you’re currently working on
Recently I finished two of my technical papers based on the techniques to compress deep neural networks. In the first paper, I proposed a system model which motivates the use of On-Device AI and Blockchain for facilitating the supply chain in agriculture. In my second project, I proposed a novel technique to compress state-of-the-art deep learning models and analysed their suitability for resource-constrained execution.
What areas do you see yourself working on in the future
Explainable AI has caught my attention recently, so I would certainly like to explore this domain for several applications including agriculture and healthcare. Also, it would be interesting to see how technologies like on-device AI can revolutionise the next-generation cyber-physical systems.
What have you learned/discovered during a particular project
Starting my PhD during a pandemic has been quite tough for a person like me. It was challenging initially to stay organised while working from home, but time got us through it. With the consistent support of my supervisor Dr Steven Davy, I was able to generate ideas and shape them into the form of a technical research paper.
I learnt many things while performing experiments and benchmarking using Python, Tensorflow, and Keras. Most importantly, I learnt how to analyse experimental results so they are convincing to reviewers of top journals and conferences.
The challenging and enjoyable aspects of being a researcher in your field.
Every PhD student initially struggles to maintain a work-life balance. I think the key to staying organised is to give ourselves mini-deadlines. In computer science, the enjoyable part is that we can work from anywhere at any time, as long as we have internet access. However, the challenging part is to keep on top of rapidly advancing research, especially when we are on verge of proposing a new idea.
Outline why your research is necessary for the end user: i.e. what are the benefits
Mobile and other IoT devices will soon be equipped with powerful computing platforms. Thus, it becomes imperative to shift AI-based decision-making from the cloud to the device. On-Device AI provides decisions with low latency along with privacy, as data is processed near to the user.
How will it improve the current state
Previous works rely on offloading a few or all computations to the edge/cloud for task completion. We propose a solution that avoids the computationally intensive offloading and ranking steps, by directly deploying a compressed model on the device. The model is compressed dynamically based on the available resource budget.
When will it be implemented
Initial results on benchmark datasets and models have been submitted for potential publication. At the moment, I am extending the work and analysing its applicability to more complex scenarios.
What are the real-world implications
There has been a recent surge in the use of Autonomous Systems for a variety of tasks such as monitoring the movement of livestock in a dairy farm using a drone. Such tasks are latency and privacy-sensitive, thus moving data out of the farm is not preferred. Therefore, the use of on-device AI can provide a quicker, reliable, and secure response as data is processed within the boundaries of the farm. Also, to enable on-device execution, it is important to compress the complexity of AI models as per the device’s resources.