Workshop
Scientific Methods for Understanding Neural Networks
Zahra Kadkhodaie 路 Florentin Guth 路 Sanae Lotfi 路 Davis Brown 路 Micah Goldblum 路 Valentin De Bortoli 路 Andrew Saxe
Sun 15 Dec, 8:50 a.m. PST
While deep learning continues to achieve impressive results on an ever-growing range of tasks, our understanding of the principles underlying these successes remains largely limited. This problem is usually tackled from a mathematical point of view, aiming to prove rigorous theorems about optimization or generalization errors of standard algorithms, but so far they have been limited to overly-simplified settings. The main goal of this workshop is to promote a complementary approach that is centered on the use of the scientific method, which forms hypotheses and designs controlled experiments to test them. More specifically, it focuses on empirical analyses of deep networks that can validate or falsify existing theories and assumptions, or answer questions about the success or failure of these models. This approach has been largely underexplored, but has great potential to further our understanding of deep learning and to lead to significant progress in both theory and practice. The secondary goal of this workshop is to build a community of researchers, currently scattered in several subfields, around the common goal of understanding deep learning through a scientific lens.
Schedule
Sun 8:50 a.m. - 9:00 a.m.
|
Opening Remarks
(
Opening remarks by organizers
)
>
SlidesLive Video |
馃敆 |
Sun 9:00 a.m. - 9:30 a.m.
|
Tom Goldstein: Can transformers solve harder problems than they were trained on? Scaling up test-time computation via recurrence
(
Keynote Talk
)
>
SlidesLive Video |
Tom Goldstein 馃敆 |
Sun 9:30 a.m. - 10:00 a.m.
|
Surya Ganguli: An analytic theory of creativity in convolutional diffusion models
(
Keynote Talk
)
>
SlidesLive Video |
Surya Ganguli 馃敆 |
Sun 10:00 a.m. - 10:30 a.m.
|
Hanie Sedghi: Exploring and Improving Planning Capabilities of LLMs
(
Keynote Talk
)
>
SlidesLive Video |
Hanie Sedghi 馃敆 |
Sun 10:30 a.m. - 10:50 a.m.
|
Coffee Break
(
Coffee Break
)
>
|
馃敆 |
Sun 10:50 a.m. - 11:05 a.m.
|
Christos Perivolaropoulos: Softmax is not enough (for sharp out-of-distribution)
(
Contributed Talk
)
>
SlidesLive Video |
Petar Veli膷kovi膰 路 Christos Perivolaropoulos 路 Federico Barbero 路 Razvan Pascanu 馃敆 |
Sun 11:05 a.m. - 11:20 a.m.
|
David Krueger: Input Space Mode Connectivity in Deep Neural Networks
(
Contributed Talk
)
>
SlidesLive Video |
Jakub Vrabel 路 Ori Shem Ur 路 Yaron Oz 路 David Krueger 馃敆 |
Sun 11:20 a.m. - 12:20 p.m.
|
Structured Identity Mapping Learning As a Model for Compositional Generalization in Generative Models ( Poster Session ) > link | Yongyi Yang 路 Core Francisco Park 路 Ekdeep S Lubana 路 Maya Okawa 路 Wei Hu 路 Hidenori Tanaka 馃敆 |
Sun 11:20 a.m. - 12:20 p.m.
|
Exploiting Interpretable Capabilities with Concept-Enhanced Diffusion and Prototype Networks ( Poster Session ) > link | Alba Carballo Castro 路 Sonia Laguna 路 Moritz Vandenhirtz 路 Julia Vogt 馃敆 |
Sun 11:20 a.m. - 12:20 p.m.
|
Distributional Scaling Laws for Emergent Capabilities ( Poster Session ) > link | Rosie Zhao 路 Naomi Saphra 路 Sham Kakade 馃敆 |
Sun 11:20 a.m. - 12:20 p.m.
|
Input Space Mode Connectivity in Deep Neural Networks ( Poster Session ) > link | Jakub Vrabel 路 Ori Shem Ur 路 Yaron Oz 路 David Krueger 馃敆 |
Sun 11:20 a.m. - 12:20 p.m.
|
Learned Random Label Predictions as a Neural Network Complexity Metric ( Poster Session ) > link | Marlon Becker 路 Benjamin Risse 馃敆 |
Sun 11:20 a.m. - 12:20 p.m.
|
Investigating Sensitive Directions in GPT-2: An Improved Baseline and Comparative Analysis of SAEs ( Poster Session ) > link | Daniel Lee 路 Stefan Heimersheim 馃敆 |
Sun 11:20 a.m. - 12:20 p.m.
|
BatchTopK Sparse Autoencoders ( Poster Session ) > link | Bart Bussmann 路 Patrick Leask 路 Neel Nanda 馃敆 |
Sun 11:20 a.m. - 12:20 p.m.
|
On the Collapse Errors Induced by the Deterministic Sampler for Diffusion Models ( Poster Session ) > link | Zhang 路 Difan Zou 馃敆 |
Sun 11:20 a.m. - 12:20 p.m.
|
softmax is not enough (for sharp out-of-distribution) ( Poster Session ) > link | Petar Veli膷kovi膰 路 Christos Perivolaropoulos 路 Federico Barbero 路 Razvan Pascanu 馃敆 |
Sun 11:20 a.m. - 12:20 p.m.
|
Understanding Visual Concepts Across Models ( Poster Session ) > link | Brandon Trabucco 路 Max Gurinas 路 Kyle Doherty 路 Ruslan Salakhutdinov 馃敆 |
Sun 11:20 a.m. - 12:20 p.m.
|
Eliminating Position Bias of Language Models: A Mechanistic Approach ( Poster Session ) > link | Ziqi Wang 路 Hanlin Zhang 路 Xiner Li 路 Kuan-Hao Huang 路 Chi Han 路 Shuiwang Ji 路 Sham Kakade 路 Hao Peng 路 Heng Ji 馃敆 |
Sun 11:20 a.m. - 12:20 p.m.
|
Interpretability as Compression: Reconsidering SAE Explanations of Neural Activations ( Poster Session ) > link | Kola Ayonrinde 路 Michael Pearce 路 Lee Sharkey 馃敆 |
Sun 11:20 a.m. - 12:20 p.m.
|
Sparse autoencoders for dense text embeddings reveal hierarchical feature sub-structure ( Poster Session ) > link | Christine Ye 路 Charles O'Neill 路 John Wu 路 Kartheik Iyer 馃敆 |
Sun 11:20 a.m. - 12:20 p.m.
|
Learnability in the Context of Neural Tangent Kernels ( Poster Session ) > link | Progyan Das 路 Dwip Dalal 馃敆 |
Sun 11:20 a.m. - 12:20 p.m.
|
The Master Key Filters Hypothesis: Deep Filters Are General ( Poster Session ) > link | Zahra Babaiee 路 Peyman M. Kiasari 路 Daniela Rus 路 Radu Grosu 馃敆 |
Sun 11:20 a.m. - 12:20 p.m.
|
The Unreasonable Ineffectiveness of the Deeper Layers ( Poster Session ) > link | Andrey Gromov 路 Kushal Tirumala 路 Hassan Shapourian 路 Paolo Glorioso 路 Dan Roberts 馃敆 |
Sun 11:20 a.m. - 12:20 p.m.
|
Probing the Decision Boundaries of In-context Learning in Large Language Models Download PDF ( Poster Session ) > link | Siyan Zhao 路 Tung Nguyen 路 Aditya Grover 馃敆 |
Sun 11:20 a.m. - 12:20 p.m.
|
The Pitfalls of Memorization: When Memorization Hinders Generalization ( Poster Session ) > link | Reza Bayat 路 Mohammad Pezeshki 路 Elvis Dohmatob 路 David Lopez-Paz 路 Pascal Vincent 馃敆 |
Sun 11:20 a.m. - 12:20 p.m.
|
"Twin Studies" of Factors in OOD Generalization ( Poster Session ) > link | Victoria R. Li 路 Jenny Kaufmann 路 David Alvarez-Melis 路 Naomi Saphra 馃敆 |
Sun 11:20 a.m. - 12:20 p.m.
|
Hiding in a Plain Sight: Out-of-Distribution Data in the Logit Space Embeddings ( Poster Session ) > link | Vangjush Komini 路 Sarunas Girdzijauskas 馃敆 |
Sun 11:20 a.m. - 12:20 p.m.
|
Is Saliency Really Captured By Gradient? ( Poster Session ) > link | Nehal Yasin 路 Jonathon Hare 路 Antonia Marcu 馃敆 |
Sun 11:20 a.m. - 12:20 p.m.
|
A Continuous-Time Analysis of Adaptive Optimization and Normalization ( Poster Session ) > link | Rhys Gould 路 Hidenori Tanaka 馃敆 |
Sun 11:20 a.m. - 12:20 p.m.
|
Logicbreaks: A Framework for Understanding Subversion of Rule-based Inference ( Poster Session ) > link | Anton Xue 路 Avishree Khare 路 Rajeev Alur 路 Surbhi Goel 路 Eric Wong 馃敆 |
Sun 11:20 a.m. - 12:20 p.m.
|
Revealing the Learning Process in Reinforcement Learning Agents Through Attention-Oriented Metrics ( Poster Session ) > link | Charlotte Beylier 路 Simon M. Hofmann 路 Nico Scherf 馃敆 |
Sun 11:20 a.m. - 12:20 p.m.
|
Training Neural Networks for Modularity aids Interpretability ( Poster Session ) > link | Satvik Golechha 路 Dylan Cope 路 Nandi Schoots 馃敆 |
Sun 11:20 a.m. - 12:20 p.m.
|
Denoising for Manifold Extrapolation ( Poster Session ) > link | Zeyu Yun 路 Galen Chuang 路 Derek Dong 路 Yubei Chen 馃敆 |
Sun 11:20 a.m. - 12:20 p.m.
|
A Method on Searching Better Activation Functions ( Poster Session ) > link | Haoyuan Sun 路 Zihao Wu 路 Bo Xia 路 Pu Chang 路 Zibin Dong 路 Yifu Yuan 路 Yongzhe Chang 路 Xueqian Wang 馃敆 |
Sun 11:20 a.m. - 12:20 p.m.
|
Pre-processing and Compression: Understanding Hidden Representation Refinement Across Imaging Domains via Intrinsic Dimension ( Poster Session ) > link | Nicholas Konz 路 Maciej Mazurowski 馃敆 |
Sun 11:20 a.m. - 12:20 p.m.
|
Model Recycling: Model component reuse to promote in-context learning ( Poster Session ) > link | Lindsay Smith 路 Chase Goddard 路 Vudtiwat Ngampruetikorn 路 David Schwab 馃敆 |
Sun 11:20 a.m. - 12:20 p.m.
|
Emergence of Hierarchical Emotion Representations in Large Language Models ( Poster Session ) > link | Bo Zhao 路 Maya Okawa 路 Eric Bigelow 路 Rose Yu 路 Tomer Ullman 路 Hidenori Tanaka 馃敆 |
Sun 11:20 a.m. - 12:20 p.m.
|
Towards Understanding In-Context Learning with Contrastive Demonstrations and Saliency Maps ( Poster Session ) > link | Fuxiao Liu 馃敆 |
Sun 11:20 a.m. - 12:20 p.m.
|
Amplified Early Stopping Bias: Overestimated Performance with Deep Learning ( Poster Session ) > link | Nona Rajabi 路 Antonio Ribeiro 路 Miguel Vasco 路 Danica Kragic 馃敆 |
Sun 11:20 a.m. - 12:20 p.m.
|
The Evolution of Statistical Induction Heads: In-Context Learning Markov Chains ( Poster Session ) > link | Ezra Edelman 路 Nikolaos Tsilivis 路 Surbhi Goel 路 Benjamin Edelman 路 Eran Malach 馃敆 |
Sun 11:20 a.m. - 12:20 p.m.
|
Transformers can reinforcement learn to approximate Gittins Index ( Poster Session ) > link | Vladimir Petrov 路 Nikhil Vyas 路 Lucas Janson 馃敆 |
Sun 11:20 a.m. - 12:20 p.m.
|
Understanding the Limitations of B-Spline KANs: Convergence Dynamics and Computational Efficiency ( Poster Session ) > link | Avik Pal 路 Dipankar Das 馃敆 |
Sun 11:20 a.m. - 12:20 p.m.
|
Structure Development in List Sorting Transformers ( Poster Session ) > link | Einar Urdshals 路 Jasmina Urdshals 馃敆 |
Sun 11:20 a.m. - 12:20 p.m.
|
Robust Learning in Bayesian Parallel Branching Graph Neural Networks: The Narrow Width Limit ( Poster Session ) > link | Zechen Zhang 路 Haim Sompolinsky 馃敆 |
Sun 11:20 a.m. - 12:20 p.m.
|
Testing knowledge distillation theories with dataset size ( Poster Session ) > link | Giulia Lanzillotta 路 Felix Sarnthein 路 Gil Kur 路 Thomas Hofmann 路 Bobby He 馃敆 |
Sun 12:20 p.m. - 1:20 p.m.
|
Lunch break
(
Lunch break
)
>
|
馃敆 |
Sun 1:20 p.m. - 1:35 p.m.
|
Antonio Sclocchi: Unraveling the Latent Hierarchical Structure of Language and Images via Diffusion Models
(
Contributed Talk
)
>
SlidesLive Video |
Antonio Sclocchi 路 Noam Levi 路 Alessandro Favero 路 Matthieu Wyart 馃敆 |
Sun 1:35 p.m. - 1:50 p.m.
|
Bao Pham: Memorization to Generalization: The Emergence of Diffusion Models from Associative Memory
(
Contributed Talk
)
>
SlidesLive Video |
Bao Pham 路 Gabriel Raya 路 Matteo Negri 路 Mohammed Zaki 路 Luca Ambrogioni 路 Dmitry Krotov 馃敆 |
Sun 1:50 p.m. - 2:20 p.m.
|
Zico Kolter: Is this really science? A lukewarm defense of alchemy
(
Keynote Talk
)
>
SlidesLive Video |
J. Zico Kolter 馃敆 |
Sun 2:20 p.m. - 2:50 p.m.
|
Misha Belkin: Building on observations: some personal experience
(
Keynote Talk
)
>
SlidesLive Video |
Misha Belkin 馃敆 |
Sun 2:50 p.m. - 3:10 p.m.
|
Coffee Break
(
Coffee Break
)
>
|
馃敆 |
Sun 3:10 p.m. - 4:10 p.m.
|
Yasaman Bahri, Andrew Gordon Wilson, Misha Belkin, Tom Goldstein, Eero Simoncelli
(
Panel discussion
)
>
SlidesLive Video |
Yasaman Bahri 路 Andrew Wilson 路 Misha Belkin 路 Eero Simoncelli 路 Tom Goldstein 路 Surya Ganguli 馃敆 |
Sun 4:10 p.m. - 4:30 p.m.
|
Winners Announcement + Closing Remarks
(
Winners Announcement + Closing Remarks
)
>
SlidesLive Video |
馃敆 |
Sun 4:30 p.m. - 5:30 p.m.
|
Memorization to Generalization: The Emergence of Diffusion Models from Associative Memory ( Poster Session ) > link | Bao Pham 路 Gabriel Raya 路 Matteo Negri 路 Mohammed Zaki 路 Luca Ambrogioni 路 Dmitry Krotov 馃敆 |
Sun 4:30 p.m. - 5:30 p.m.
|
Language model scaling laws and zero-sum learning ( Poster Session ) > link | Andrei Mircea 路 Ekaterina Lobacheva 路 Supriyo Chakraborty 路 Nima Chitsazan 路 Irina Rish 馃敆 |
Sun 4:30 p.m. - 5:30 p.m.
|
Evaluating Loss Landscapes from a Topology Perspective ( Poster Session ) > link | Tiankai Xie 路 Caleb Geniesse 路 Jiaqing Chen 路 Yaoqing Yang 路 Dmitriy Morozov 路 Michael Mahoney 路 Ross Maciejewski 路 Gunther Weber 馃敆 |
Sun 4:30 p.m. - 5:30 p.m.
|
Training Dynamics of Convolutional Neural Networks for Learning the Derivative Operator ( Poster Session ) > link | Erik Wang 路 Yongji Wang 路 Ching-Yao Lai 馃敆 |
Sun 4:30 p.m. - 5:30 p.m.
|
Understanding the Transient Nature of In-Context Learning: The Window of Generalization ( Poster Session ) > link | Core Francisco Park 路 Ekdeep S Lubana 路 Hidenori Tanaka 馃敆 |
Sun 4:30 p.m. - 5:30 p.m.
|
Learning Stochastic Rainbow Networks ( Poster Session ) > link | Vivian White 路 Muawiz Chaudhary 路 Guy Wolf 路 Guillaume Lajoie 路 Kameron Decker Harris 馃敆 |
Sun 4:30 p.m. - 5:30 p.m.
|
SolidMark: How to Evaluate Memorization in Image Generative Models ( Poster Session ) > link | Nicky Kriplani 路 Minh Pham 路 Malikka Rajshahi 路 Chinmay Hegde 路 Niv Cohen 馃敆 |
Sun 4:30 p.m. - 5:30 p.m.
|
Token-token correlations predict the scaling of the test loss with the number of input tokens ( Poster Session ) > link | Francesco Cagnetta 路 Matthieu Wyart 馃敆 |
Sun 4:30 p.m. - 5:30 p.m.
|
Explicit Regularisation, Sharpness and Calibration ( Poster Session ) > link | Israel Mason-Williams 路 Fredrik Ekholm 路 Ferenc Huszar 馃敆 |
Sun 4:30 p.m. - 5:30 p.m.
|
Knowledge Distillation for Teaching Symmetry Invariances ( Poster Session ) > link | Patrick Odagiu 路 Nicole Nobili 路 Fabian Dionys Schrag 路 Yves Bicker 路 Yuhui Ding 馃敆 |
Sun 4:30 p.m. - 5:30 p.m.
|
Standard adversarial attacks only fool the final layer ( Poster Session ) > link | Stanislav Fort 馃敆 |
Sun 4:30 p.m. - 5:30 p.m.
|
Generalization vs Specialization under Concept Shift ( Poster Session ) > link | Alex Nguyen 路 David Schwab 路 Vudtiwat Ngampruetikorn 馃敆 |
Sun 4:30 p.m. - 5:30 p.m.
|
Unraveling the Latent Hierarchical Structure of Language and Images via Diffusion Models ( Poster Session ) > link | Antonio Sclocchi 路 Noam Levi 路 Alessandro Favero 路 Matthieu Wyart 馃敆 |
Sun 4:30 p.m. - 5:30 p.m.
|
Stitching Sparse Autoencoders of Different Sizes ( Poster Session ) > link | Patrick Leask 路 Bart Bussmann 路 Joseph Bloom 路 Curt Tigges 路 Noura Al Moubayed 路 Neel Nanda 馃敆 |
Sun 4:30 p.m. - 5:30 p.m.
|
Illusions as features: the generative side of recognition ( Poster Session ) > link | Tahereh Toosi 路 Kenneth Miller 馃敆 |
Sun 4:30 p.m. - 5:30 p.m.
|
Exploring model depth and data complexity through the lens of cellular automata ( Poster Session ) > link | Tianyu He 路 Darshil Doshi 路 Aritra Das 路 Andrey Gromov 馃敆 |
Sun 4:30 p.m. - 5:30 p.m.
|
We Need Far Fewer Unique Filters Than We Thought ( Poster Session ) > link | Zahra Babaiee 路 Peyman M. Kiasari 路 Daniela Rus 路 Radu Grosu 馃敆 |
Sun 4:30 p.m. - 5:30 p.m.
|
How Learning Rates Shape Neural Network Focus: Insights from Example Ranking ( Poster Session ) > link | Ekaterina Lobacheva 路 Keller Jordan 路 Aristide Baratin 路 Nicolas Le Roux 馃敆 |
Sun 4:30 p.m. - 5:30 p.m.
|
Alice in Wonderland: Simple Tasks Reveal Severe Generalization and Basic Reasoning Deficits in State-Of-the-Art Large Language Models ( Poster Session ) > link | Marianna Nezhurina 路 Lucia Cipolina Kun 路 Mehdi Cherti 路 Jenia Jitsev 馃敆 |
Sun 4:30 p.m. - 5:30 p.m.
|
Causation Does Not Imply Correlation: A Study of Circuit Mechanisms and Model Behaviors ( Poster Session ) > link | Jenny Kaufmann 路 Victoria R. Li 路 Martin Wattenberg 路 David Alvarez-Melis 路 Naomi Saphra 馃敆 |
Sun 4:30 p.m. - 5:30 p.m.
|
Effectiveness of Sparse Autoencoder for understanding and removing gender bias in LLMs ( Poster Session ) > link | Praveen Hegde 馃敆 |
Sun 4:30 p.m. - 5:30 p.m.
|
Boundaries of stable regions in activation space of LLMs become sharper with more compute ( Poster Session ) > link | Jett Janiak 路 Jacek Karwowski 路 Chatrik Mangat 路 Giorgi Giglemiani 路 Nora Petrova 路 Stefan Heimersheim 馃敆 |
Sun 4:30 p.m. - 5:30 p.m.
|
Measuring the Reliability of Causal Probing Methods: Tradeoffs, Limitations, and the Plight of Nullifying Interventions ( Poster Session ) > link | Marc Canby 路 Adam Davies 路 Chirag Rastogi 路 Julia C Hockenmaier 馃敆 |
Sun 4:30 p.m. - 5:30 p.m.
|
Impact of Label Noise on Learning Complex Features ( Poster Session ) > link | Rahul Vashisht 路 P Kumar 路 Harsha Vardhan Govind 路 Harish Guruprasad Ramaswamy 馃敆 |
Sun 4:30 p.m. - 5:30 p.m.
|
How rare events shape the learning curves of hierarchical data ( Poster Session ) > link | Hyunmo Kang 路 Francesco Cagnetta 路 Matthieu Wyart 馃敆 |
Sun 4:30 p.m. - 5:30 p.m.
|
Specialization-generalization transition in in-context learning of linear functions ( Poster Session ) > link | Chase Goddard 路 Lindsay Smith 路 Vudtiwat Ngampruetikorn 路 David Schwab 馃敆 |
Sun 4:30 p.m. - 5:30 p.m.
|
Knowledge Distillation: The Functional Perspective ( Poster Session ) > link | Gabryel Mason-Williams 路 Israel Mason-Williams 路 Mark Sandler 馃敆 |
Sun 4:30 p.m. - 5:30 p.m.
|
Are Capsule Networks Texture or Shape Biased? ( Poster Session ) > link | Riccardo Renzulli 路 Dominik Vranay 路 Marco Grangetto 馃敆 |
Sun 4:30 p.m. - 5:30 p.m.
|
Sometimes I am a Tree: Data Drives Fragile Hierarchical Generalization ( Poster Session ) > link | Tian Qin 路 Naomi Saphra 路 David Alvarez-Melis 馃敆 |
Sun 4:30 p.m. - 5:30 p.m.
|
EmoCAM: Toward Understanding What Drives CNN-based Emotion Recognition ( Poster Session ) > link | Youssef Doulfoukar 路 Laurent Mertens 路 Joost Vennekens 馃敆 |
Sun 4:30 p.m. - 5:30 p.m.
|
Improving Deep Learning Speed and Performance through Synaptic Neural Balance ( Poster Session ) > link | Antonios Alexos 路 ian domingo 路 Pierre Baldi 馃敆 |
Sun 4:30 p.m. - 5:30 p.m.
|
Is Expressivity Essential for the Predictive Performance of Graph Neural Networks? ( Poster Session ) > link | Fabian Jogl 路 Pascal Welke 路 Thomas G盲rtner 馃敆 |
Sun 4:30 p.m. - 5:30 p.m.
|
Is network fragmentation a useful complexity measure? ( Poster Session ) > link | Coenraad Mouton 路 Randle Rabe 路 Dani毛l Haasbroek 路 Marthinus Theunissen 路 Hermanus Potgieter 路 Marelie Davel 馃敆 |
Sun 4:30 p.m. - 5:30 p.m.
|
Comparing Apples and Oranges: is Stitching Similarity a Load of Spheres? ( Poster Session ) > link | Damian Smith 路 Antonia Marcu 馃敆 |
Sun 4:30 p.m. - 5:30 p.m.
|
Emergent properties with repeated examples ( Poster Session ) > link | Francois Charton 路 Julia Kempe 馃敆 |
Sun 4:30 p.m. - 5:30 p.m.
|
Rethinking Knowledge Transfer in Learning Using Privileged Information ( Poster Session ) > link | Danil Provodin 路 Bram van den Akker 路 Christina Katsimerou 路 Maurits Clemens Kaptein 路 Mykola Pechenizkiy 馃敆 |