Image Retrieval using Short Texts
summary
ImageSuggest sits between user text and external image search APIs, extracting ranked keywords from Whisper posts and feeding cached third‑party queries; it filters images by text detection, white background and resolution. Trained on Whisper logs with a fixed vocabulary of 36,000 terms, a cosine/nearest‑neighbor baseline scores ~2% accuracy, a feedforward NN reaches 13% after three days on GPU, and a bidirectional LSTM peaks at 17% after seven days. Suggestions lift creation rates above 50%.
AI-written summary · about
chapters (10)
- 0:00Introduction to talk
- 1:22Whisper platform overview
- 3:12Image suggestions importance
- 5:46Initial image retrieval challenges
- 7:14Learning from user searches
- 9:42Scalable term-based approach
- 11:14Classification model development
- 13:22LSTM model improvements
- 15:44Model comparison examples
- 16:05Q&A and hiring announcement
related talks
30mSemantic Search and Neural Information Retrieval
43mThe Search for a New Visual Search, Beyond Language
26mMaximizing Multimodal: Exploring the search frontier of text-to-image models
42mThe Machine Understands Text — Learning and Search at Micro and Macro Scale
26mImage search without image captions
05mFine-tuned SigLIP Image
33mMachine Learning in Visual Search
50mElasticsearch Suggesters: Beyond Autocomplete