Kshitij Gupta
Papers
1
Total Citations
15
H-Index
1
About
Kshitij Gupta is a researcher whose work sits at the intersection of deep learning theory and practical model evaluation. His most recognized contribution is the introduction of "Broken Neural Scaling Laws" (BNSL), a 2022 paper that has garnered 15 citations and offers a novel, smoothly broken power law functional form. This framework provides a more nuanced and accurate way to model and extrapolate the scaling behaviors of deep neural networks—specifically, how evaluation metrics shift as the compute budget for training or inference changes. By moving beyond simple power law assumptions, Gupta’s work helps practitioners better predict model performance at scale, a critical insight for resource allocation and architecture design. His research is particularly valuable for students and engineers navigating the complexities of large-scale model development, offering a practical tool for understanding when and why scaling benefits diminish. With a focus on empirical rigor and theoretical clarity, Gupta’s contributions are shaping how the community thinks about the limits and possibilities of neural network scaling.
Research Focus
Key Achievements
Top Papers
- 1Broken Neural Scaling Laws15 citations · 2022