Pragmatic Deep Learning for image labelling
summary
An image pipeline labels hotel photos using transfer learning from pre-trained models in the Caffe model zoo and public datasets (Places and SUN), then extracts feature vectors, reduces them with NMF, and plugs them into a logistic-regression meta-model that nightly scores 18M users × upcoming sales on Spark/Hadoop. A/B testing shows the production combo of image features plus content/behavior models increases revenue by +7%, and the team deploys dynamic image selection to surface higher-converting visuals.
AI-written summary · about
chapters (13)
- 0:00Introduction
- 1:50The use case
- 5:34Retargeting
- 6:35Results
- 7:47Why image labelling
- 8:55Image similarity
- 9:50Common issues
- 10:37What can we do
- 12:19Transfer learning
- 13:15Machine learning models
- 14:00Classification problem
- 16:09Classification results
- 17:55Conclusion
related talks
33mDeep Learning at Booking.com
40mHow to Implement a Production Ready Deep Learning Model in 12 Weeks
17mLaying the foundation for deep learning
46mML Use Cases at Think Big Analytics
33mRunning Machine Learning in Production - a Journey to Success
29mLabeling Tools are Great, but What About Quality Checks?
35mLevel Up Your Machine Learning Lifecycle
39mInsights About Places with Deep Learning Computer Vision