Using Computer Vision to Understand Big Visual Data
summary
Shows that modern pipelines train convolutional neural networks on millions of images to map pixels to labels and beat humans on parts of the ImageNet thousand‑class benchmark, but the systems still spit generic captions for rare cases. Claims that deep learning scales well for common classes yet fails on the long tail of uncommon visual phenomena, demonstrating working classifiers and captioners that impress on averages but break badly on surprising or specific inputs.
AI-written summary · about
chapters (6)
- 0:00Start
- 0:10Visual similarity via labels
- 0:29Machine Learning as data association
- 1:22At test time...
- 1:53Automatic Image Captioning
- 2:11Challenge 1: Visual world has Long Tails
related talks
01h 18mComputer Vision after the Victory of Data
42mDeep Learning for Computer Vision
47mDeep Learning in Computer Vision
42mState of the Art of Innovation in Computer Vision
26mComputer Vision in the Era of Deep Learning
03mRethinking software engineering in the AI era
41mDeep Learning to Solve Challenging Problems
01h 02mTutorial on Deep Learning I