EMU Seminars

Department of Mathematics, Uppsala University


Emerging Mathematica Uppsala : a seminar series on AI, learning theory, verification and computer related methods. These should be understandable to a broad mathematical audience. Talks will be Mondays 13.15-14.00 unless otherwise noted in Ångström Laboratory, Uppsala.

Organiser: Fiona Skerman. Please contact me to be added to a mailing list.


Upcoming Seminars

Mon Nov 16 13:15 - 14:00 Room Å64119 -- map (also on zoom, email for details).

AI for math: what is happening and how the community is responding.

Speaker: Emily Riehl (John Hopkins)

Abstract:

This talk will survey some notable recent applications of generative AI to research mathematics and then discuss how the community is responding.

Update : presentation slides from the talk can be found here.

Mon Nov 30 13:15 - 14:00 Room Å64119 -- map.

Speaker: Benny Avelin (UU)
Mon Dec 7 13:15 - 14:00 Room Å64119 -- map.

Speaker: Peter LeFanu Lumsdaine (SU)




Past Seminars

Mon Oct 19 13:15 - 14:00 Room Å101127 -- map.

Neuroalgebraic Geometry

Speaker: Kathlén Kohn (KTH)

Abstract:

The space of functions parametrized by a fixed neural network architecture is known as its "neuromanifold", a term coined by Amari. Training the network means to solve an optimization problem over the neuromanifold. Thus, a complete understanding of its intricate geometry would shed light on the mysteries of deep learning.

This talk explores the approach to approximate neural networks by algebraic ones that have semialgebraic neuromanifolds. Such approximation is possible for any continuous network on a compact data domain. By the universal approximation theorem, algebraic neural networks are essentially the only ones whose neuromanifolds span finite-dimensional ambient spaces. In this setting, we can interpret training the network as finding a "closest" point on the neuromanifold to some data point in the ambient space. This perspective enables us to understand the loss landscape better, which is the graph of the loss function over the neuromanifold. In particular, the singularities (and boundary points) of the neuromanifold can cause a tradeoff between efficient optimization and good generalization: On the one hand, singularities can yield numerical instability and slow the learning process (which was already observed by Amari). On the other hand, we will observe how the same singularities cause implicit bias to stable and sparse solutions.

Computing the singularities is often a technical endeavor, and requires us to determine both the hidden parameter symmetries of the network and the critical points of the network's parametrization map. This talk looks at 3 popular architectures: multilayer perceptrons, convolutional networks, and self-attention networks, and will provide many exercises!

Mon Oct 12 13:15 - 14:00 Room Å4001 -- map.

Maths formalisation in the age of AI: What the working mathematician should know about Lean.

Speaker: Yaël Dillies (SU)

Abstract:

A decade ago, a couple of mathematicians entreprised to rewrite their corpus in the extremely pedantic Lean programming language to ensure superhuman correctness of mathematical arguments and as a continuation of the Bourbaki philosophy. Now, AI is spitting mathematical breakthroughs on a daily basis, faster than humans can comprehend them.

In this talk, I will survey the ecosystem built around Lean, in particular by the aforementioned pioneer mathematicians, and explain how it will become a key component in the new age of LLM-powered mathematics that lies ahead of us. I will teach a few concrete skills on how to verify for yourself that a purported Lean proof of a theorem you care about is correct, and encourage the audience to bring their laptops to follow along (Unix-based preferably).

Mon Sep 28 13:15 - 14:00 Room Å64119 -- map.

How do I Lean? An amateur's guide to LLM-assisted formalisation.

Speaker: Sascha Troscheit (UU)

Abstract:

While I had started (and given up) on learning Lean some years ago, the rise of LLMs prompted me to reevaluate my life's choices. Since the beginning of this year, I have extensively experimented with bringing LLMs into my research. First, to aid with structuring of larger proofs, and later to formalise results in Lean. These led to three (human-created) research outcomes that I vibe-coded into Lean formalised proofs. In this seminar, I will give:
1) a brief account of what I learnt;
2) methods to structure proofs prior to formalisation and to avoid hallucinations;
3) and formalisation using LLM agents.
Bringing laptops highly encouraged to type along.

Update : presentation slides from the talk can be found here.

Mon Sep 14 13:15 - 14:00 Room Å100155 hus 10 -- map.

What is Lean? A short practical introduction.

Speaker: Julian Külshammer (UU)

Abstract:

It is now almost every day that we read a headline stating that this or that conjecture has been resolved with the help of AI. Often this comes with a second sentence: "The result has been verified in Lean." I will give a short overview of what Lean is, how to use it (mostly without AI), and some ideas what to be careful about when reading such statements. Nothing is assumed, if you want to program along, it is helpful to bring your own Laptop. There is an online platform https://live.lean-lang.org/ that can be used without installation, but installation instructions can also be found on https://lean-lang.org/install/.

Update : a link to the lean examples and presentation slides from the talk can be found here.

Resources for learning lean:

Lean Game Server (link) : lean tutorials in the browser (i.e. does not require installing lean). Natural Number Game and Reintroduction to Proofs both on the lean game server.

Formalising Mathematics course by Bhavik Mehta (link).

Mathematics in Lean (link).

Lean in browser (link).