Pick a problem with a data source attached
The most common way a student project dies is that the idea is good and the data does not exist. Before you commit, answer one question: where does the first hundred labelled examples come from, and can you get them yourself in two weeks?
If the answer is that you will collect them, that is a strong project rather than a weak one. The mushroom work in this thread started with a team photographing infected and healthy oyster mushrooms in a real farm, because no public dataset of Trichoderma infection existed. That dataset is now published openly and is the part of the project other people can build on. It also won the paper its award.
Baseline before you reach for the biggest model
Every team wants to start with the largest architecture they have heard of. Start with the smallest thing that could work and measure it, because that number is what makes the rest of the project legible. If a simple classifier gets you to seventy per cent, the contribution of the complicated one is the difference, not the whole.
It also protects you from the most common failure in a student result: a number that looks excellent because the test set leaked, or because one class dominates. If your accuracy is far above your baseline on the first attempt, something is usually wrong with the split.
- Sentiment analysis on movie reviews: a comparative analysis2023 International Conference on Intelligent Systems for Communication, IoT, 2023
- Fake Review Detection using Deep LearningArtificial Intelligence and Communication Technologies, pp. 655-668, 2023
- Fake news detection using supervised learning techniquesInternational Conference on Applied Mechanics, Machine Learning and Advanced, 2023
Measure something beyond accuracy
The projects that turn into publishable work are almost always the ones that measure a second thing. What does inference cost? How long does it take on the device it will actually run on? What does it consume?
That is the whole point of the Eco-Transformers work: the accuracy of a large language model is well studied and its energy cost per inference is not, so the contribution is in the second measurement. The same instinct drove the on-device categorisation work, where the constraint was the phone, and the solar work, where the constraint was cloud cover in real time rather than in a clean dataset.
- Eco-Transformers: Carbon-Efficient and Accelerated Inference through Self-Pruning and Activation Freezing in Large Language Models2025 ICNGCS, pp. 1-8, 2025
- On-Device Application Categorization using Application Resources2023 IEEE International Conference on Electronics, Computing and, 2023
- AI-Driven Solar Panel Optimization during Cloudy times Using Real-Time Image Processing2025 International Conference on Next Generation Computing Systems (ICNGCS), 2025
Write it up as you go
A project written up in the last fortnight reads like one. Keep a running record of what you tried and what it scored, including the attempts that failed, because the failed attempts are what make a comparison section honest and they are impossible to reconstruct from memory afterwards.
If you would like to work on something in this direction, the student path on the work-with-me page says what to send. A message that names a paper here and the question it left you with will get a real reply.
What to read next
Outside work, not mine. Each title links to a Google Scholar search, which will find you whichever version is open access.