BEGIN:VCALENDAR
VERSION:2.0
PRODID:researchseminars.org
CALSCALE:GREGORIAN
X-WR-CALNAME:researchseminars.org
BEGIN:VEVENT
SUMMARY:Fabian Ruehle (Northeastern University)
DTSTART:20251204T190000Z
DTEND:20251204T200000Z
DTSTAMP:20260404T111322Z
UID:ai-for-science/1
DESCRIPTION:Title: Symbolic Regression\, Sparsification\, and Kolmogorov-Arnold
Networks\nby Fabian Ruehle (Northeastern University) as part of First
Principles Talks\n\n\nAbstract\nInterpreting neural networks remains chall
enging\, largely due to their dense parametrization\, global coupling of p
arameters\, and the polysemantic behavior of neurons. These problems are a
meliorated in Kolmogorov-Arnold Networks\, which have fewer parameters ove
rall\, parameter changes are contained to local regions\, and there are le
ss polysemantic neurons. \n\nIn the first part of this talk\, Fabian will
show how KANs can be viewed as neural networks that have undergone a princ
ipled sparsification\, clarifying why they exhibit improved interpretabili
ty and parameter efficiency. He will then present a new framework for mult
ivariate symbolic regression that couples KANs\, LLMs\, and genetic search
strategies\, akin to FunSearch\, to discover compact analytic expressions
from data. This approach enables scalable symbolic regression in high-dim
ensional settings\, leverages the inductive biases inherent in KANs\, and
the ability to prime the LLM's regression proposals for different data dom
ains.\n
LOCATION:https://stable.researchseminars.org/talk/ai-for-science/1/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Juan Felipe Carrasquilla Alvarez (ETH Zurich)
DTSTART:20251211T150000Z
DTEND:20251211T160000Z
DTSTAMP:20260404T111322Z
UID:ai-for-science/2
DESCRIPTION:Title: Language Models for the Simulation of Quantum Many-Body\
nby Juan Felipe Carrasquilla Alvarez (ETH Zurich) as part of FirstPrincipl
es Talks\n\n\nAbstract\nIn this talk\, Juan will discuss his work on using
models inspired by natural language processing in the realm of quantum ma
ny-body physics. He will demonstrate their utility in solving ground state
s of quantum Hamiltonians\, particularly for ground states of arrays of Ry
dberg atoms on the Kagome lattice. The findings highlight the potential of
using language models to explore many-body physics on exotic lattices and
beyond.\n
LOCATION:https://stable.researchseminars.org/talk/ai-for-science/2/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Jianke Yang (UC San Diego)
DTSTART:20251218T190000Z
DTEND:20251218T200000Z
DTSTAMP:20260404T111322Z
UID:ai-for-science/3
DESCRIPTION:Title: Towards AI Co-Scientist: Automatic Governing Law Discovery\nby Jianke Yang (UC San Diego) as part of FirstPrinciples Talks\n\n\nAb
stract\nScientific law discovery has historically been limited by human re
asoning and data scarcity\, despite the vast search space of possible form
ulations. Advances in generative AI and abundant physical data now enable
AI models to extract interpretable structures\, such as symmetries\, diffe
rential equations\, and conserved quantities\, and use them as inductive b
iases in predictive and generative tasks.\n\n\nJianke's research aims to d
evelop an AI co-scientist\, a unified system that can (1) autonomously dis
cover governing structures from raw observations\, (2) translate these dis
coveries into flexible inductive biases to improve downstream tasks\, and
(3) orchestrate modular tools under a top-level planner to generate hypoth
eses\, implement models that satisfy physical constraints\, and complete t
he pipeline from data→law→model→prediction.\n\n\nThis thesis proposa
l presents the following milestones toward this goal. First\, we formulate
the problem of symmetry discovery and introduce two models\, LieGAN and L
aLiGAN\, that discover invariance and equivariance from data using a gener
ative adversarial framework. Second\, we incorporate symmetry into the tas
k of governing equation discovery\, showing that symmetry is a powerful in
ductive bias in the discovery of other physical laws. Together\, these ser
ve as the building blocks towards a fully functional AI co-scientist syste
m.\n
LOCATION:https://stable.researchseminars.org/talk/ai-for-science/3/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Gerard Milburn (National Quantum Computing Centre)
DTSTART:20260108T160000Z
DTEND:20260108T170000Z
DTSTAMP:20260404T111322Z
UID:ai-for-science/4
DESCRIPTION:Title: Quantum machines learning quantum\nby Gerard Milburn (Na
tional Quantum Computing Centre) as part of FirstPrinciples Talks\n\nAbstr
act: TBA\n
LOCATION:https://stable.researchseminars.org/talk/ai-for-science/4/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Nathan Kutz (University of Washington)
DTSTART:20260113T200000Z
DTEND:20260113T210000Z
DTSTAMP:20260404T111322Z
UID:ai-for-science/5
DESCRIPTION:Title: Shallow Recurrent Decoders for the Automated Discovery of Ph
ysical Models\nby Nathan Kutz (University of Washington) as part of Fi
rstPrinciples Talks\n\nAbstract: TBA\n
LOCATION:https://stable.researchseminars.org/talk/ai-for-science/5/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Chenhao Tan (University of Chicago)
DTSTART:20260116T203000Z
DTEND:20260116T213000Z
DTSTAMP:20260404T111322Z
UID:ai-for-science/6
DESCRIPTION:Title: Science in the Age of AI\nby Chenhao Tan (University of
Chicago) as part of FirstPrinciples Talks\n\nAbstract: TBA\n
LOCATION:https://stable.researchseminars.org/talk/ai-for-science/6/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Alexei Koulakov (Cold Spring Harbor Laboratory)
DTSTART:20260128T180000Z
DTEND:20260128T190000Z
DTSTAMP:20260404T111322Z
UID:ai-for-science/7
DESCRIPTION:Title: From Neurons to Newtons: What can the brain teach us about p
hysics?\nby Alexei Koulakov (Cold Spring Harbor Laboratory) as part of
FirstPrinciples Talks\n\n\nAbstract\nModern physics has been extraordinar
ily successful at describing the natural world\, yet the process by which
new physical theories are constructed remains largely artisanal. In this t
alk\, Alexei will discuss the principles of brain function and evolution w
hich can offer tools for building new physics theories. \n\nFirst\, he wil
l introduce the concept of a genomic bottleneck\, the idea that neural sys
tems are forced to compress vast sensory experience into representations t
hat are simple\, robust\, and reusable across tasks. I suggest that simila
r bottlenecks may be essential for identifying abstractions that generaliz
e across subfields of physics. Second\, he will discuss how brains appear
to construct internal imagination modules\, generative models that allow o
rganisms to simulate physical phenomena and test hypotheses without direct
interaction with the world. Finally\, Alexei will show how hierarchical r
einforcement learning can provide a natural framework for organizing physi
cal reasoning across scales\, from low-level dynamics to high-level concep
ts. \n\nBy decomposing complex problems into nested objectives\, hierarchi
cal control offers a computational model for how intelligent systems\, bio
logical or artificial\, can efficiently explore and solve hard physics pro
blems. These ideas suggest a neuroscience-inspired roadmap for transformin
g theory building in physics: one that emphasizes distillation\, imaginati
on\, and hierarchical control as core computational primitives.\n
LOCATION:https://stable.researchseminars.org/talk/ai-for-science/7/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Randy Davilla (Rice University\; FirstPrinciples)
DTSTART:20260331T150000Z
DTEND:20260331T160000Z
DTSTAMP:20260404T111322Z
UID:ai-for-science/8
DESCRIPTION:Title: Automated Conjecturing\, Internal Theories\, and AI Discover
y\nby Randy Davilla (Rice University\; FirstPrinciples) as part of Fir
stPrinciples Talks\n\n\nAbstract\nAutomated conjecturing is the computer-a
ssisted generation of human-readable mathematical statements—such as bou
nds and structural patterns—that experts can test\, refine\, or prove. B
uilding on earlier systems like Graffiti and TxGraffiti\, this approach fo
cuses on producing interpretable conjectures rather than predictive models
.\n\nIn this talk\, we introduce Graffiti3\, a framework that organizes ma
thematical evidence into evolving tables of objects and invariants\, allow
ing candidate conjectures to be generated\, tested\, and refined through c
ounterexample search and expert review. The system combines deterministic
conjecture engines with large language models to propose features and summ
arize results\, illustrating how AI systems can assist in mathematical dis
covery across domains such as graph theory\, finite groups\, knot theory\,
and Calabi–Yau geometry.\n
LOCATION:https://stable.researchseminars.org/talk/ai-for-science/8/
END:VEVENT
BEGIN:VEVENT
SUMMARY:Philip Harris (MIT)
DTSTART:20260416T180000Z
DTEND:20260416T190000Z
DTSTAMP:20260404T111322Z
UID:ai-for-science/9
DESCRIPTION:Title: Around the Forces in 80 Microseconds\nby Philip Harris (
MIT) as part of FirstPrinciples Talks\n\n\nAbstract\nWith large amounts of
data\, a Higgs boson discovery\, and world-leading constraints on fundame
ntal forces\, the Large Hadron Collider has been a phenomenal tool. Howeve
r\, it is going through a midlife crisis. More data\, more Higgs bosons\,
and more constraints are not generating the same excitement as in the past
. We venture into a new direction with fresh insights that enable unpreced
ented physics measurements\, and we ask how we can automate the discovery
and measurement process using artificial intelligence. We then look at the
future of the LHC and present a real-time system\, a custom electronics s
ystem\, built around similar novel AI-based processing technology that wil
l expand the scope of future physics measurements at the LHC. We extend th
e same real-time AI approaches into gravitational wave astrophysics\, high
lighting new results with an end-to-end AI pipeline. Finally\, we comment
on a new paradigm for next-generation experimental physics research.\n
LOCATION:https://stable.researchseminars.org/talk/ai-for-science/9/
END:VEVENT
END:VCALENDAR