I have worked on prior projects which might be insightful for readers, or just provide potentially interesting ideas for people to work on. Some of them are detailed below.

Industry Work
	NVIDIA Hopper GPU: Scaling Performance NVIDIA Corp. Hot Chips 34, 2022 A technical deep dive into the Hopper GPU and its features. I specifically developed hardware for fast dependency resolution and cluster grids, as detailed on pages 22-26.
Technical Reports
	Automatic fine-tuning Nipun Agarwala, Haider Suleman, Neal Vaidya CS 330, 2022 We investigate strategies to automatically finetune large pre-trained deep learning models with few shot learning.
	Finding Sparse Approximate Matrix Inverses Nipun Agarwala, Kao Kitichotkul, Shaked Regev EE 346B, 2021 Given an sparse matrix, we try to find efficiently find a highly accurate sparse approximate inverse without full Gaussian Elimination.
	Music Composition using Recurrent Neural Networks Yuki Inoue, Nipun Agarwala, Axel Sly CS 224N, 2017 We investigate generative RNN models that create sheet music with well-formed structure and stylistic conventions without predefining music composition rules to the models.
	End-to-end neural networks for subvocal speech recognition Pol Rosello, Pamela Toman, Nipun Agarwala CS 224S, 2017 We describe an end-to-end, session-independent subvocal speech recognition by leveraging character-level recurrent neural networks despite limited data constraints.
	CGRA Architecture Tian Zhao, Nipun Agarwala, David Koeplinger, Yaqi Zhang, Artem Vasilyev, Raghu Prabhakar CS 315B, 2016 We try to find a general CGRA architecture suitable for parallel patterns, including machine learning.
	Hitting depth: Investigating robustness to adversarial examples in deep convolutional neural networks Chris Billovits, Mihail Eric, Nipun Agarwala CS 231N, 2016 We study the nature of adversarial examples and their impact on robustness of deep models, and then investigate a few methods to mitigate them in neural networks.
Other Work
	A case for managed and model-less inference serving Neeraja J Yadwadkar, Francisco Romero, Qian Li, Christos Kozyrakis, HOT OS, 2019 The authors make a case for automated and model-less inferencing to serve the large number of applications using servers for inferencing (My role was more supportive and consulting).
	Optimizing SRAM sizing using Convex Methods Kevin Kiningham, Nipun Agarwala EE 213, 2016 A fun little project where we tried to optimize SRAM sizings using convex optimization methods!