Scholar
Sina Alemohammad
Google Scholar ID: ATjmZVsAAAAJ
Postdoctoral Fellow, University of Texas at Austin
Citations & Impact
Citations
428
H-index
5
i10-index
3
Publications
15
Co-authors
8
list available
Publications
Resume (English only)
Academic Achievements
- Proposed the Model Autophagy Disorder (MAD) framework, revealing degradation in generative models trained on their own synthetic data.
- Developed Neon: a post-hoc method that fine-tunes on synthetic data and reverses degradation, achieving SOTA FID of 1.02 on ImageNet-256 with only 0.36% extra compute.
- Introduced SIMS: a training framework using a model’s own synthetic data as negative guidance, setting new FID records on CIFAR-10 and ImageNet-64 while preventing model collapse.
- Co-developed WaLRUS: integrates wavelet transforms with state-space models for effective long-range sequence modeling.
- Co-proposed SaFARi: a frame-agnostic extension of state-space models supporting arbitrary functional bases for flexible long-range representation.
- Contributed to TITAN: enhances implicit neural representations by integrating deep image priors via a residual deep decoder.
- Research featured in major outlets including The New York Times, New Scientist, Fortune, Sciencedaily, France 24, and Times of India.
Background
- Currently a postdoctoral fellow in the VITA Group at the University of Texas at Austin, working under the mentorship of Prof. Atlas Wang.
- Research focuses on the theory of deep learning and the development of generative models.
- Introduced the concept of Model Autophagy Disorder (MAD), a phenomenon where models degrade in realism and diversity due to repeated training on their own synthetic outputs.
- Current work aims to understand and prevent MAD, developing strategies for models to self-improve with synthetic data without falling into self-consuming feedback loops.
- Broader research interests include deep learning theory, generative modeling, and sparse signal processing.
