Machine Learning in Visual Search

on YouTube ↗33m

" summary

Visual search drives higher conversions and is practical today using either image-to-text tagging stored in Elastic or true image-to-image search via VGG16 feature extraction and nearest-neighbor vectors. A simple CNN on Fashion-MNIST hits about 90% accuracy in tests, but real-world problems—posture and background bias, dominant-color mistakes, and noisy user photos—require massive, relevant training sets, frequent retraining (eBay retrains every few days), and sub-second response times (1–5s load causes ~90% abandonment). Use Google Cloud or Clarifai to prototype fast.

AI-written summary · about

" chapters (12)

1. 0:00Start
2. 2:36Why We Even Need Visual Search and Why It Blooms
3. 6:48Why Visual Search Was Around for Years
4. 9:47Image to Text or Image To Image
5. 18:10Sample Project
6. 22:28Challenges What's Hard in Visual Search
7. 23:44.Training Data versus Real-Life
8. 24:33Data Set Relevancy
9. 25:21Colors
10. 27:30Training Data Size
11. 28:00Response Time
12. 31:47Questions

" related talks

• 28mVisual SearchMarcin Tuszyński · PyData Warsaw 2019
• 42mDeep Learning for Computer VisionAlex Conway · PyCon.DE 2017
• 43mThe Search for a New Visual Search, Beyond LanguageBryan Minor · Strange Loop 2018
• 41mSearching for Products by PhotoIvan Krasnikov · Saint HighLoad++ 2023
• 31mAI Search for Marketplaces: Experiences & TechMladen Fernezir · DSC Europe 24
• 22mHarnessing image and text data for product matchingThijs Roukens, Gokhan Simsek · Devworld Conference 2025
• 22mMachine Learning: The Engine Behind Modern E-commerce SearchMilutin Studen · DSC Europe 24
• 30mEvolving Search in Online Marketplaces: Practical Applications of AIMladen Fernižer · Shift 2025