Sairam S
Bengaluru, Karnataka, India 560078.
I am a third-year undergraduate researcher at PES University. My research focuses on Controlled Failure Modes in Deep Learning currently looking at geometric instability in Low-Precision Optimization.
Some of my projects:
- The Geometry of Bit-Collapse Investigating the spectral impact of Quantization. Demonstrated that Int8 quantization induces sharp minima (Bit-Collapse) and that SAM restores geometric robustness.
- First-Principles Engineering: To master the mechanics of deep learning, I engineered a complete, reverse-mode automatic differentiation engine in Python from the ground up. I then built and trained a neural network using only this custom framework, demonstrating a ground-up understanding of gradient-based learning.
news
| Jan 01, 2026 | Completed preliminary research on “The Geometry of Bit-Collapse,” identifying a 5.7x spectral spike in the Hessian of Int8 quantized networks. Visit the Repository |
|---|---|
| Nov 25, 2025 | Submitted first-author paper, “Systematic Dissection of Graph R-Learners,” to Transactions on Machine Learning Research (TMLR). (Under Review). Paper |
| Aug 12, 2025 | 🎓 Selected as Teaching Assistant for Mathematics for Computer Science at PES University. |