The TWIML AI Podcast (formerly This Week in Machine Learning & Artificial Intelligence) Managing Deep Learning Experiments with Lukas Biewald - TWIML Talk #295
Aug 29, 2019
Lukas Biewald, CEO and co-founder of Weights & Biases, dives into the world of deep learning experiments. He shares how his company addresses the need for reproducibility and effective evaluation frameworks. The conversation covers the chaos of machine learning management and the importance of collaboration in improving workflows. Lukas also teases an upcoming ebook on data management and highlights the significance of version control in community-driven projects. Learn how his tools are reshaping the experimentation landscape in AI!
AI Snips
Chapters
Transcript
Episode notes
Lukas's Career Path
- Lukas Biewald started at Yahoo, working on machine learning ranking systems.
- He later worked at PowerSet, a natural language search company, before founding Figure Eight.
Experiment Tracking Chaos
- Sam Charrington mentions seeing hyperparameters embedded in file names.
- Lukas Biewald recalls using a single, large Emacs file with sporadic notes for experiment tracking in grad school.
Reproducibility Crisis
- Reproducing experiments is difficult due to factors beyond code, like hyperparameters and datasets.
- Recreating past experiments for papers can be challenging if not tracked meticulously.