Guidelines for proposing your own project for B.Sc. students

This page serves as a guideline for B.Sc. students interested in proposing their own project. If you are a M.Sc. student, these rules do not apply to you.

Your thesis work should, as a generic guideline, be an original and novel piece of research (although there might be exceptions, scroll down for more details). This means that, before presenting your idea for a thesis, you should ensure that your idea is not already present in the literature. This is achieved via an accurate literature review, which is a time-consuming process that is usually one of the first processes which you, as a thesis student, are required to do by the University guidelines. This work should not be done wholely by the supervisor. This means that, if you want to propose a project, you should already demonstrate some knowledge of the field achieved via a proper literature review.

To showcase your proposal, I suggest you prepare a short presentation clearly stating (a) the problem you want to tackle, (b) relevant works already done in the field, and (c) your original idea for tackling the project.

Note that very common topics, such as neural networks to predict stock markets or the outcome of football matches, or other cliché computer vision problems, are connected to an extremely large amount of literature: the chance of someone else having already considered your approach is hence very high. In addition, a literature review might be close to intractable in these cases. I highly suggest you avoid these topics.

When is it OK to propose non-novel research?

There are some cases in which it is OK to propose a non-novel research project. These cases are mainly connected to reproducibility issues:

• Some code might not be available, hence you may propose to implement methods described in a paper.
• Some other papers might release their code, but the results might not be reproducible: bugs happen, and a lot of papers fail to release essential information for reproducibility (hyperparameters, specifics of the datasets, etc.).

Notice that you are still required to make the case for a non-novel project: the motivation should be good enough to justify the work, and the re-implementation or the experimental design must still be challenging enough for a thesis.

In other cases, instead, you might be interested in benchmarking multiple methods on a common ground, such that the results are comparable. If you intend on pursuing this road, you should also ensure that no pre-existing benchmarks exist, or that you want to do something substantially different than other benchmarks.

Areas I might consider for a supervision on a project proposal

My main fields of interest/research are:

• Model compression
• Trustworthy AI (and, more specifically, XAI and fairness)
• Deep Learning for Computer Vision (bonus point if connected to applications of digital humanities)

I might also consider projects connected to Deep Learning for audio data.

Areas I will not consider for a supervision on a project proposal

I will not consider projects connected to:

• Natural Language Processing: we already have multiple experts on that field in our department (I’m not one of them), you should refer to them
• Reinforcement Learning: see above
• Time Series Analysis: using Deep Learning for time series analysis is a very challenging topic on which I have little expertise; moreover, the application of Deep Learning to, e.g., financial data, is often a cliché with an exterminate amount of existing literature and very little chance for a successful novel approach.
• Datasets with unlabelled data unless you intend to use semi-/self-/un-supervised learning: the process of data labelling is time-consuming and often requires a lot of “unskilled” manual work, which is not a proper way to use your time for a B.Sc. thesis.