Finding needles in haystacks with Deep Neural Networks
summary
Claims a learned manifold pulls duplicate product images close and different products far using a multi‑task classifier as a base, then metric learning with contrastive loss and later triplet loss. Uses Alex‑style conv outputs (6×6×256) precomputed, small fully connected metric network; cleans noisy catalog with perceptual hashing to collapse ~10M products to ~1M reliable groups. Training runs take many hours and need fixes like gradient clipping and skipping L2 normalization; results are mixed — roughly half of sampled queries match well, work in progress.
AI-written summary · about
chapters (21)
- 0:00Neural network journey
- 1:08Duplicate products problem
- 3:36Product image perspectives
- 4:37Simple classifier for poses
- 7:26Pose clustering in feature space
- 8:49Metric learning with symmetric networks
- 10:13Contrastive loss training results
- 11:58Data noise and duplicate cleaning
- 12:42Perceptual hashing for images
- 14:06Triplet loss implementation struggles
- 16:30Effectiveness of data
- 17:13Debugging two-product test case
- 18:16Gradient clipping for stability
- 19:20Network learning in 2D
- 20:21Data sampling issues
- 21:28L2 normalization debate
- 23:15Final working model
- 25:03Early results and convergence
- 25:46Model predictions and limits
- 26:32Key lessons learned
- 29:01Q&A
related talks
34mComputer Vision and NLP for Multi-Task Fashion Attrb. Modeling
22mHarnessing image and text data for product matching
28mVisual Search
41mSearching for Products by Photo
33mMachine Learning in Visual Search
11mComparing apples to apple trees: Injecting domain knowledge to similarity searches
36mNear-Duplicate Ad Detection in Online Classified Ad Services
33mDeep Learning at Booking.com