Neural Collisions

No upcoming talks at the moment..

• 04.06.2026 Thursday @ 16:00 CET

Speaker:     Vasilis Niarchos (University of Crete)

Title: CFTs are lazy 

Abstract:     How much information is needed to determine a correlation function in conformal field theory? Surprisingly, very little. I will present a novel approach based on neural networks that determines correlation functions in arbitrary CFTs in any spacetime dimension from almost nothing. I will discuss the underlying mechanisms behind this approach and its potential wide-ranging implications.

Slides

• 21.05.2026 Thursday @ 16:00 CET

Speaker:     Mao Zeng (University of Edinburgh)

Title: Machine Learning for Optimizing Feynman Integral Algorithms

Abstract:     Precision calculations in quantum field theory are crucial for collider and gravitational wave physics but often face a major computational bottleneck: reducing large families of Feynman integrals to a small set of master integrals. This reduction is based on systematic integration-by-parts (IBP) identities, but making it fast in practice requires many heuristic algorithmic choices. Small changes in these choices can have an enormous impact on performance. In this talk, I will discuss how machine learning can help guide these choices. In earlier work, we used reinforcement learning and metaheuristics to search automatically for efficient IBP reduction strategies. In ongoing work, these machine-discovered patterns lead to simple, human-understandable strategies: using sparse, carefully chosen sets of seed integrals and IBP equations that still contain enough information to complete the reduction. In the examples we study, this turns otherwise very expensive high-rank reductions into problems whose cost grows only linearly with the numerator power. I will illustrate this with high-rank nonplanar integrals that are very difficult to reduce using conventional strategies, and I will also review and comment on related work.

Slides

• 07.05.2026 Thursday @ 16:00 CET

Speaker:     Jesse Thaler (MIT/IAIFI)

Title: "Descending into the Modular Bootstrap"

Abstract:     The modular bootstrap has been a powerful tool for carving out the landscape of allowed two-dimensional conformal field theories (CFTs).  In this talk, I describe a complementary approach to standard modular bootstrap bounds: using modern machine learning strategies to actively search for CFT spectra that yield a valid torus partition function.  Using insights from statistical inference and a custom singular-value-based optimizer, I present evidence for an obstruction to finding CFTs with small central charge and large spectral gaps, and I speculate on what this might imply for the structure of the CFT landscape.  Along the way, I reflect on "centaur" approaches to theoretical physics, where human physicists and artificial intelligence collaborate to explore spaces of theories that would be difficult to navigate alone.

Slides

• 23.04.2026 Thursday @ 16:00 CET

Speaker:   G. Bruno de Luca (IFT Madrid)

Title: Energy Conserving Descent and applications 

Abstract:     Training can be viewed as dynamics on a loss landscape. Standard optimization methods are friction dominated: they dissipate energy and may stop in poor local minima. Energy Conserving Descent replaces this picture with Hamiltonian dynamics, where convergence arises through kinetic effects, conserving a natural notion of energy. This yields a family of optimizers that can be analyzed with tools such as Liouville’s theorem: when the dynamics mixes, one obtains a calculable distribution over parameters and hence predictive control over optimization outcomes. I will outline the framework, describe concrete algorithms derived from it and ways to promote mixing, and discuss applications ranging from precision particle physics inference to PDE solving. The same construction also turns naturally into sampling, yielding MCHMC and its variants, which have already shown promise in applications such as cosmological field-level inference and Bayesian neural network sampling. 

Based on: 2501.18419, 2306.00352, 2212.08549, 2201.11137.

Slides

• 19.03.2026 Thursday @ 16:00 CET

Speaker:     Fabian Ruehle (Northeastern/IAIFI)

Title: "Kolmogorov Arnold Networks and Differential equations"

Abstract:     In this talk I will introduce Kolmogorov Arnold Networks as an alternative Neural Network architecture. I will describe the theory behind KANs and their properties and argue that KANs are particularly well-suited for symbolic regression and hence for scientific discovery. I will also review their applications to PDEs. 

Slides