Research

Empirical Papers

In addition to theoretical work, empirical studies are important to advance any scientific field. However, it is always difficult to pose the right questions or define meaningful hypotheses, to design experiment that answer the questions or validate hypotheses, and to discuss the observations. Here, we collected good empirical works in ML and Systems community.

• ARE ALL LAYERS CREATED EQUAL?
• Cancellation in Systems: An Empirical Study of Task Cancellation Patterns and Failures
• The Effects of Regularization and Data Augmentation are Class Dependent
• MLaaS in the Wild
• Measuring Robustness to Natural Distribution Shifts in Image Classification
• Transfer Learning for Performance Modeling of Configurable Systems: An Exploratory Analysis
• Statically Inferring Performance Properties of Software Configurations
• Do vision transformers see like convolutional neural networks?
• Pretrained transformers as universal computation engines
• Evaluation of neural architectures trained with square loss vs cross-entropy in classification tasks
• Model evaluation, model selection, and algorithm selection in machine learning
• Multi-tenant GPU Clusters for Deep Learning Workloads: Analysis and Implications
• Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers
• EfficientNetV2: Smaller Models and Faster Training
• Why Do Better Loss Functions Lead to Less Transferable Features?
• Do Better ImageNet Models Transfer Better?

SoK Papers

SoK papers are important.

• SoK in security
• SoK in JSys

Vision Papers