Data, Science: Investigating 1 Million Galaxies with Humans and TensorFlow
summary
Combines crowdsourced labels and TensorFlow to analyze roughly 1 million galaxies using volunteers treated as classifiers with per-user confusion matrices, calibrated by simulated lenses, and aggregated with Bayesian updating; uses ensembles / Bayesian neural networks at test time to estimate uncertainty and drive selective labeling. Human+model ensembles outperform either alone, enable novel discoveries, and scale labeling, while failure modes include sparse/conflicting labels, artifact-filled images, and decision-tree questions that break confusion-matrix assumptions.
AI-written summary · about
chapters (13)
- 0:00Introduction
- 2:59Crowdsourcing science
- 4:02Reliable scientific labels
- 10:03Volunteer performance
- 14:39Machine learning models
- 16:25Recap of convolutional neural networks
- 21:11Human-machine predictions
- 24:20Uncertain data selection
- 27:18Uncertainty estimation
- 29:28Human-in-the-loop training
- 30:52Unexpected discoveries
- 32:20Human feedback importance
- 33:01Conclusion and call to action
related talks
37mCounting the Stars
41mExploring Dark Energy & Star Formation w/ ML
01h 00mData, visualization, and designing AI
41mModels Schm-odels: Why you Should Care about Data-Centric AI
19mIdentifying Personal Data by Fusing Elasticsearch with Neural Networks
29mCritical CV-NLP Data Errors and How to Fix Them with Galileo
04mUsing Computer Vision to Understand Big Visual Data
51mEfficient Deep Learning with Humans in the Loop