The TWIML AI Podcast (formerly This Week in Machine Learning & Artificial Intelligence) The Whys and Hows of Managing Machine Learning Artifacts with Lukas Biewald - #373
12 snips
May 7, 2020 Lukas Biewald, Founder and CEO of Weights & Biases, dives into the world of machine learning artifact management. He shares insights about their new tool, Artifacts, designed to track datasets, models, and pipelines seamlessly. The conversation also highlights the challenges of data provenance and reproducibility, alongside the evolution from simplistic methods to more advanced solutions. Biewald emphasizes the importance of user-friendly approaches and the need for organizations to understand their data processes for successful integration and project outcomes.
AI Snips
Chapters
Transcript
Episode notes
Artifacts: A New ML Tool
- Weights & Biases' new Artifacts tool tracks datasets, models, and pipelines, similar to experiment tracking.
- It versions and saves these artifacts, or pointers to them, addressing the need for reproducibility and lineage tracking.
Evolving Datasets in Retail and Robotics
- Retail companies training models for automatic product detection face constantly evolving datasets.
- Privacy takedowns, like those experienced by iRobot, further complicate dataset consistency.
Data Selection and Preprocessing
- Autonomous vehicle companies, with vast datasets, selectively train models on specific data subsets.
- This necessitates meticulous tracking, especially of preprocessing steps for medical companies with limited, valuable data.