DataEngBytes 2023 - SYD-T2-04 - Suneeta Mall
summary
Argues for treating data as the development driver, not an afterthought, and shows concrete techniques: automated label-error detection using cleanlab and confident learning to find mislabeled runway/airplane annotations, empirical triaging into three-sample buckets (easy once, critical repeat, never-learns) that cuts redundant training and compute, and a UDA-style iterative loop that measures veracity, value, and volume so teams reduce storage and labeling cost while preserving model performance.
AI-written summary · about
chapters (8)
- 0:00Introduction
- 4:56Talk motivation
- 7:46Iterative data development
- 12:22Seven data dimensions in AI
- 14:08Data veracity at scale
- 16:38Maximizing training data value
- 19:08Optimizing data volume
- 20:30Q&A
related talks
33mDataEngBytes 2023 - MEL-T1-02 - Kendra Vant
27mDataEngBytes 2023 - PER-T1-08 - Sergio Zenatti Filho & Elizabeth Antoinne
28mDataEngBytes 2023 - MEL-T1-11 - Emily Melhuish
03mDBTB INT Jill Desmond
39mDataEngBytes 2023 - BRI-T1-02 - Mikiko Bazeley
28mDataEngBytes 2023 - PER-T1-05 - Ellie Gholami
24mDataEngBytes 2023 - MEL-T1-05 - Izzudin Hafiz
24mDataEngBytes 2023 - MEL-T2-02 - Vikas Rajput
