Alex Damian

Timeline

1. 2016

   An Angier B. Duke scholarship

   A graduate of Devon Preparatory School in Pennsylvania, he receives an Angier B. Duke Memorial Scholarship for undergraduate study at Duke University in 2016.

2. 2020

   PULSE generates detailed faces

   At Duke University, he coauthors PULSE in 2020 as part of Cynthia Rudin's team. The system searches images made by a generative adversarial network to create detailed faces that match a low-resolution input when reduced in size.

3. 2022

   Learning representations with gradient descent

   With Jason Lee and Mahdi Soltanolkotabi, he coauthors Neural Networks can Learn Representations with Gradient Descent at the Conference on Learning Theory in 2022. For problems with low-dimensional structure, the work shows how gradient descent trains a two-layer neural network to learn representations relevant to its task.

4. 2023

   MeZO for language models

   He coauthors Fine-Tuning Language Models with Just Forward Passes at NeurIPS in 2023. The team introduces MeZO, an optimizer that fine-tunes language models using forward passes with the same memory footprint as inference.

5. 2024

   Transformers learn causal structure

   With Eshaan Nichani and Jason D. Lee, he coauthors How Transformers Learn Causal Structure with Gradient Descent at ICML in 2024. On a designed learning task, they prove that a simplified two-layer transformer trained with gradient descent learns to encode a hidden causal graph in its first attention layer.

6. 2025

   A doctorate at Princeton

   After a BS in mathematics at Duke University, he completes a PhD in Princeton's Program in Applied and Computational Mathematics, supervised by Jason D. Lee. His final oral public examination in June 2025 presents Foundations of Deep Learning: Optimization and Representation Learning, a thesis on representation learning and optimization dynamics.

7. 2025

   Kempner research fellow at Harvard

   In 2025, he becomes a Kempner Research Fellow at the Kempner Institute for the Study of Natural and Artificial Intelligence at Harvard University. He studies how training methods, including stochastic gradient descent and Adam, shape the representations learned by neural networks.

8. 2026

   An AI2050 fellowship

   In 2026, he is named an AI2050 Early Career Fellow. His project develops a mathematical theory of how optimizers interact with neural network architectures, with the aim of deriving effective training procedures automatically.

9. Since 2026

   Assistant professor at MIT

   In 2026, he joins MIT as an assistant professor with a shared appointment in the Departments of Mathematics and Electrical Engineering and Computer Science through the Schwarzman College of Computing. He is also a member of LIDS and studies the mathematical foundations of deep learning, including optimization dynamics and representation learning.