1 of 12

Team Presentation

MAchine Learning @ poliTO

2 of 12

About

Us

3 of 12

Context

  • How does a challenge work?
    • There is a public leaderboard where you can make submissions and get a score. At the end of the challenge period, the final ranking will be drawn up from here!

​

  • Machine learning challenges:
    • Conferences (NeurIPS, KDD, CIKM)
    • Workshops (SemEval, EVALITA)
    • Industry-sponsored (Amazon, Google, IBM)
    • Community-based (Kaggle, DataDriven)
    • Hackathons

4 of 12

Malto – Our Goals

  • Creating the first Data Science & ML student team�at PoliTO, with the goal of:
    • building an environment for data science enthusiasts
    • developing technical skills in the field of ML through hands-on projects
    • gain experience working in teams
    • supervision by PhD and PoliTO researchers with expertise in the field
    • promote the visibility of PoliTO locally, nationally and internationally

​

  • Participating to international competitions
    • conferences and challenges held by universities and companies

5 of 12

Past

Achievements

6 of 12

Our experience - Sigspatialcup 2023

  • Task: Supraglacial Lake Satellite Image Segmentation
  • Approach :
    • Segmentation of images using CNNs
    • Post-processing generating polygons of recognized lake from model’s prediction
  • 6th place
  • Won the $1,000 USD travel award to go to the Sigspatial conference in Hamburg in November 2023

7 of 12

Neurips 2024 - Machine Unlearning

  • Task: Machine Unlearning

​

  • The objective of machine unlearning is to remove the influence of selected information from a model. Information in fact can be easily removed from any database, whereas removing it from a model can be much more complicated.

​

  • 12th place in public leaderboard. Scientific paper on the work done under review in the 39th Annual AAAI Conference

https://www.kaggle.com/competitions/neurips-2023-machine-unlearning/discussion/458648

8 of 12

SemEval 2024 - Shroom

  • Task: Detect LLM hallucinations

​

  • LLMs like ChatGPT-4 are known to hallucinate information. A response is hallucinated when the generated text is confident and syntactically correct, but the information, for various reasons, is totally made up and therefore wrong.

​

  • 13th place + publication of a paper describing the approach adopted (https://aclanthology.org/2024.semeval-1.240.pdf)

9 of 12

ECML PKDD 2024 - VOLVO Discovery Challenge

  • Task:

Develop a model to predict the risk levels to help Volvo find the components requiring maintenance. The provided dataset contains information about two generations of the same component.

​

  • Subtasks:
    • Accurately predict the risk level of a component from the same generation as the provided training data
    • Model that performs well on the second generation (generalization)
  • 3rd place and live presentation of our approach at the conference!

10 of 12

SemEval 2025- Machine Unlearning LLM

  • Task: Machine Unlearning

​

  • The objective of machine unlearning is to remove the sensitive content from a large language models . Information in fact can be easily removed from any database, whereas removing it from a model can be much more complicated.

​

  • 8th place in public leaderboard. Scientific paper on the work done under review in the ACL 2025

https://aclanthology.org/2025.semeval-1.229.pdf

11 of 12

Workshop

📅 4 Oct – Data Exploration & Classification� 📅 11 Oct – Regression & Deep Learning� 📍 Room 29 @ PoliTo

💻 Hands-on coding, demos, Q&A� 🔷 Open to all (basic Python 🐍 needed)� 🙌 Boost skills & MALTO recruitment chances

​

​

12 of 12

Join us!

Follow us on Instagram for the future events and recruitment moments of the team!!!