Data Science Talks
The Data Science Talks are a recurring event of the Data Science Initiative. They are designed to facilitate the exchange of ideas and inspiration. The event consists of several snappy (5-7 min) presentations, followed by a casual gathering with refreshments:
1. Washing away Adam's limitations by Urh Prosenc (Data Scientist, Teads)
Adam has been the default optimizer for deep learning, but its diagonal preconditioner struggles on complex, non-isotropic loss landscapes. In this session, we look at SOAP—a next-generation optimizer that brings full second-order preconditioning down to a practical scale. We look at how SOAP is able to smooth out steep gradient descent bottle necks, outpace Adam's convergence and deliver squeeky-clean performance gains without blowing up compute budgets.
2. Is Technology Making Us Dumber? by Erik Štrumbelj (Researcher, UL FRI)
Digital technology is deeply embedded in our everyday lives, with AI-based tools taking this integration even further. But what does this mean for how we think, learn, remember, and pay attention? I’ll give a quick tour of what research tells us about the potential negative effects of digital technology on cognition, highlighting the main cognitive domains affected, where the evidence is strongest, and where important uncertainties remain.
3. From AutoML to AgenticML: LLMs as the New SearchOperator by Sebastian Koralewski (Data Scientist,Teads)
Every automated search process - tuning hyperparameters, evolving code, running a research pipeline — follows the same loop: propose a candidate → run it → read the result → decide what to try next. With the rise of LLMs, the question is no longer whether this loop can be automated, but which parts of it already can be and where a human still needs to stay in it.
This talk introduces a five-level ladder that sorts systems by how much of the loop has been handed from human to model:
- A0 — Manual. Human runs the whole loop.
- A1 — Propose. Model suggests the next candidate. Human still owns the program, the runs, the search policy.
- A2 — Reflect. Model turns written feedback into a suggested edit. Human still writes and runs the program.
- A3 — Author. Model writes the program.
- A4 — Research. Model runs the research loop itself, picks the question and drafts the write-up.
4. Everyone Uses AI ... But How? Insights from Future Software Engineers by Damjan Fujs (Assistant Professor, UL FRI)
Generative AI is reshaping software development and software engineering education. Drawing on experiences from project-based teaching and recent research on student teams, this talk explores how AI changes the entire development lifecycle, from requirements engineering to deployment. Using findings on the gap between students’ perceived usefulness of LLM and their actual usage, I discuss emerging developer skills, the future of development teams, and challenges for education and industry.
5. Snacks and drinks and discussions