Export
Rudy Chin
· 2 previous affiliations
- Carnegie Mellon UniversityCurrent
- Current
- PhD student · 2017–2022
- MS student · 2015–2017
Research Overview
Rudy Chin focuses on neural network optimization, model pruning, and the intersection of machine learning with hardware architecture. His trajectory spans techniques for reducing the computational overhead of deep learning models, including channel optimization and slimmable networks, alongside investigations into the hyperparameter dynamics of magnitude pruning. Most recently, his work addresses pedagogical strategies for teaching machine learning to engineering students.
Research Trajectory
In 2020, Chin led research on optimizing the computational efficiency of convolutional neural networks (CNNs). As first author on Width transfer: on the (in)variance of width optimization, he investigated CNN width optimization, a technique used to adjust channel counts across network layers to maintain accuracy while minimizing floating-point operations (FLOPs) during inference. This work utilized sensitivity analysis to address the computational overhead typically associated with treating width optimization as a hyperparameter search, aiming to significantly speed up the training process.
During the same period, Chin extended this focus to resource-constrained environments, such as mobile devices, through his first-author work on PareCO: Pareto-aware Channel Optimization for Slimmable Neural Networks. This research addressed limitations in standard slimmable neural networks, which traditionally applied a uniform width-multiplier across all layers. Chin's work highlighted that different layers have varying impacts on prediction accuracy and computational requirements, necessitating a more principled approach to assigning layer-specific width-multipliers to achieve optimal performance trade-offs.
Also in 2020, Chin contributed as second author to Uncovering the impact of hyperparameters for global magnitude pruning. This research examined unstructured magnitude pruning and the lottery ticket framework in vision classification tasks. The study challenged the implicit assumption that the best training configuration for evaluating a model is also optimal for discovering its sparse mask. By decoupling the hyperparameters used for mask discovery from those used for mask evaluation, the authors identified a phenomenon where sub-optimal evaluation configurations could actually yield masks that ultimately train to higher performance.
In 2021, Chin's published work transitioned toward engineering education. As first author on Putting the "Machine" Back in Machine Learning for Engineering Students, he explored effective teaching strategies for integrating computer hardware architecture into deep learning coursework, highlighting the critical role hardware plays in advancing machine learning capabilities.
Current Directions
The most recent paper in the evidence is from 2021 and indicates an interest in the pedagogy of machine learning. This direction focuses on the intersection of computer hardware architecture and deep learning education, addressing how engineering curricula can better incorporate the hardware fundamentals that underpin modern machine learning systems and applications.
Profile updated
| Title | Venue | Year |
|---|---|---|
| Width Transfer: On the (In)variance of Width Optimization | Workshops @ CVPR | 2021 |
| PareCO: Pareto-aware Channel Optimization for Slimmable Neural Networks | Openreview | 2020 |
| Uncovering the impact of hyperparameters for global magnitude pruning | Openreview | 2020 |
Loading new papers…
Other researchers whose papers cluster into the same research directions — a proxy for overlapping technical interests and shared intellectual neighborhoods.
See something wrong? Report a mistake