BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//linuxsoftware.nz//NONSGML Joyous v1.4//EN
BEGIN:VEVENT
SUMMARY:Bridging Theory and Machine Learning: Analytical Insights into Neu
 ral Networks and Generative Models
DTSTART:20250313T100000Z
DTEND:20250313T120000Z
DTSTAMP:20260801T110204Z
UID:c2621030-79d3-4830-9515-c9b82a4ab02f
SEQUENCE:2
CREATED:20250310T114314Z
DESCRIPTION:Modern artificial intelligence has achieved state-of-the-art p
 erformance across various domains\, from solving protein folding to predic
 ting new materials. However\, establishing a solid theoretical foundation 
 for machine learning remains an ongoing research challenge. This seminar e
 xplores how analytical models can help bridge this gap.In the first part\,
  we examine the learning behavior of neural networks near their optimal po
 int. Specifically\, we analyze the Hessian of the loss function with respe
 ct to the learnable parameters in teacher-student setups where both networ
 ks share the same architecture. By characterizing the Hessian eigenspectru
 m for different activation functions\, we show that the Hessian rank at th
 e optimal solution effectively determines the number of relevant parameter
 s. In the second part\, we discuss recent research on applying generative 
 diffusion models to classical spin systems\, such as the Ising model. The 
 Ising model serves as a testbed\, allowing us to leverage its well-underst
 ood analytical properties to explore how different design choices in diffu
 sion models impact generative performance.
LAST-MODIFIED:20250310T114339Z
LOCATION:DF Seminar Room (2-8.3)\, 2nd floor of Physics Building
URL:http://df.vps.tecnico.ulisboa.pt/pt/eventos/bridging-theory-and-machin
 e-learning-analytical-insights-into-neural-networks-and-generative-models/
X-ALT-DESC;FMTTYPE=text/html:<p data-block-key="epnco">Modern artificial i
 ntelligence has achieved state-of-the-art performance across various domai
 ns\, from solving protein folding to predicting new materials. However\, e
 stablishing a solid theoretical foundation for machine learning remains an
  ongoing research challenge. This seminar explores how analytical models c
 an help bridge this gap.</p><p data-block-key="8lm1v"><br/>In the first pa
 rt\, we examine the learning behavior of neural networks near their optima
 l point. Specifically\, we analyze the Hessian of the loss function with r
 espect to the learnable parameters in teacher-student setups where both ne
 tworks share the same architecture. By characterizing the Hessian eigenspe
 ctrum for different activation functions\, we show that the Hessian rank a
 t the optimal solution effectively determines the number of relevant param
 eters.<br/><br/> In the second part\, we discuss recent research on applyi
 ng generative diffusion models to classical spin systems\, such as the Isi
 ng model. The Ising model serves as a testbed\, allowing us to leverage it
 s well-understood analytical properties to explore how different design ch
 oices in diffusion models impact generative performance.</p>
END:VEVENT
END:VCALENDAR