Abstract:Reproducing published deep learning papers to validate their conclusions can be difficult due to sources of irreproducibility. We investigate the impact that implementation factors have on the results and how they affect reproducibility of deep learning studies. Three deep learning experiments were ran five times each on 13 different hardware environments and four different software environments. The analysis of the 780 combined results showed that there was a greater than 6% accuracy range on the same deterministic examples introduced from hardware or software environment variations alone. To account for these implementation factors, researchers should run their experiments multiple times in different hardware and software environments to verify their conclusions are not affected.
Abstract:Machine learning (ML) research has generally focused on models, while the most prominent datasets have been employed for everyday ML tasks without regard for the breadth, difficulty, and faithfulness of these datasets to the underlying problem. Neglecting the fundamental importance of datasets has caused major problems involving data cascades in real-world applications and saturation of dataset-driven criteria for model quality, hindering research growth. To solve this problem, we present DataPerf, a benchmark package for evaluating ML datasets and dataset-working algorithms. We intend it to enable the "data ratchet," in which training sets will aid in evaluating test sets on the same problems, and vice versa. Such a feedback-driven strategy will generate a virtuous loop that will accelerate development of data-centric AI. The MLCommons Association will maintain DataPerf.