Keys to Data-Centric AI: Asking AI to Improve its own Dataset.
summary
Claims that using a baseline model plus simple algorithms finds and fixes dataset problems so models improve without code changes; open source clean laab implements confident learning and flags label errors, outliers, near-duplicates and bad annotators by using embeddings and predicted class probabilities from any model. The workflow deletes or relabels a small fraction of data, yields large accuracy gains across image, text and tabular sets, and scales to LLM fine-tuning and RLHF where noisy labels cripple learning.
AI-written summary · about
chapters (9)
- 0:00Introduction to data-centric AI
- 3:00Why data quality matters
- 7:36The data engine concept
- 13:15Detecting label errors with AI
- 19:54Data-centric vs model-centric methods
- 24:27Improving LLM fine-tuning with clean data
- 27:59Multi-annotator label consensus
- 30:47Fixing noisy sequence data
- 31:29Q&A
related talks
30mData-Centric AI Cookbook-let's prep that data
41mModels Schm-odels: Why you Should Care about Data-Centric AI
11mData-centric AI: A new paradigm
37mBest Practices for Building AI Datasets
30mActive Learning and Human-in-the-Loop Machine Learning
14mData Lifecycle Management as Foundational Pillar to AI Agents
28mData Curation: Transforming Bytes into AI Gold
21mDataEngBytes 2023 - SYD-T2-04 - Suneeta Mall