Models Schm-odels: Why you Should Care about Data-Centric AI
summary
Claim: iterating on data beats fiddling with models for practical gains. Concrete actions include cleaning noisy labels, rebalancing train and validation splits, and adding counterfactuals. Tools actually used: ResNet50 baseline for embeddings, UMAP + Altair visualization to spot distribution shifts, Google Sheets for collaborative labeling, and a Streamlit drawable canvas to generate augmented examples. Constraints and surprises: a six week contest with a 10 000 example submission cap, a 3 000 image dataset, third place in a prior competition, and winning the competition’s innovative category shows the approach works.
AI-written summary · about
chapters (17)
- 0:00Data-centric AI intro
- 2:52Why data prep matters
- 3:56Lung cancer detection case
- 7:26Data-Centric AI competition
- 9:35Better labels with tools
- 12:47Manual analysis lessons
- 14:35Embeddings for data issues
- 17:05Train/validation rebalancing
- 17:26Counterfactual data creation
- 19:54Tooling for data iteration
- 20:14Top teams' approaches
- 23:48Why data-centric now
- 26:59Data quality impact
- 28:42Collaboration benefits
- 32:54Good data traits
- 36:02Shift to data-centric AI
- 38:55Q&A: Label disagreements
related talks
30mData-Centric AI Cookbook-let's prep that data
33mKeys to Data-Centric AI: Asking AI to Improve its own Dataset.
11mData-centric AI: A new paradigm
01h 00mData, visualization, and designing AI
26mUnlocking the Power of Data-Centric AI.
46mWhy data-driven methods will shape the future (...)
37mBest Practices for Building AI Datasets
17mFocus on Data, on the Road to AI