Picture for Alex Damian

Alex Damian

Understanding Optimization in Deep Learning with Central Flows

Add code
Oct 31, 2024
Viaarxiv icon

Computational-Statistical Gaps in Gaussian Single-Index Models

Add code
Mar 12, 2024
Viaarxiv icon

How Transformers Learn Causal Structure with Gradient Descent

Add code
Feb 22, 2024
Viaarxiv icon

Fine-Tuning Language Models with Just Forward Passes

Add code
May 27, 2023
Viaarxiv icon

Smoothing the Landscape Boosts the Signal for SGD: Optimal Sample Complexity for Learning Single Index Models

Add code
May 18, 2023
Viaarxiv icon

Provable Guarantees for Nonlinear Feature Learning in Three-Layer Neural Networks

Add code
May 11, 2023
Viaarxiv icon

Self-Stabilization: The Implicit Bias of Gradient Descent at the Edge of Stability

Add code
Sep 30, 2022
Figure 1 for Self-Stabilization: The Implicit Bias of Gradient Descent at the Edge of Stability
Figure 2 for Self-Stabilization: The Implicit Bias of Gradient Descent at the Edge of Stability
Figure 3 for Self-Stabilization: The Implicit Bias of Gradient Descent at the Edge of Stability
Figure 4 for Self-Stabilization: The Implicit Bias of Gradient Descent at the Edge of Stability
Viaarxiv icon

Neural Networks can Learn Representations with Gradient Descent

Add code
Jun 30, 2022
Figure 1 for Neural Networks can Learn Representations with Gradient Descent
Figure 2 for Neural Networks can Learn Representations with Gradient Descent
Viaarxiv icon

Label Noise SGD Provably Prefers Flat Global Minimizers

Add code
Jun 11, 2021
Figure 1 for Label Noise SGD Provably Prefers Flat Global Minimizers
Figure 2 for Label Noise SGD Provably Prefers Flat Global Minimizers
Figure 3 for Label Noise SGD Provably Prefers Flat Global Minimizers
Figure 4 for Label Noise SGD Provably Prefers Flat Global Minimizers
Viaarxiv icon

New Techniques for Preserving Global Structure and Denoising with Low Information Loss in Single-Image Super-Resolution

Add code
Jun 16, 2018
Figure 1 for New Techniques for Preserving Global Structure and Denoising with Low Information Loss in Single-Image Super-Resolution
Figure 2 for New Techniques for Preserving Global Structure and Denoising with Low Information Loss in Single-Image Super-Resolution
Figure 3 for New Techniques for Preserving Global Structure and Denoising with Low Information Loss in Single-Image Super-Resolution
Figure 4 for New Techniques for Preserving Global Structure and Denoising with Low Information Loss in Single-Image Super-Resolution
Viaarxiv icon