DBTB INT Maarten
summary
Claims that more data is created in the last two years than in all previous history and uses that scale to power suggesting images to posts by mapping short text to visuals. Uses a generic deep learning model rather than bespoke features, trains on large-scale noisy pairs, and evaluates retrieval accuracy and qualitative relevance; reports that the model generalizes across tasks and yields useful image suggestions in production after iterative tuning and dataset curation.
AI-written summary · about
related talks
01h 18mComputer Vision after the Victory of Data
01h 23mMultimodal Deep Learning in the Real World
30mHow to train a large diffusion model
46mWhy data-driven methods will shape the future (...)
30m🤯 No data? No problem! Synthetic data to the rescue
11mData-centric AI: A new paradigm
21mDataEngBytes 2023 - SYD-T2-04 - Suneeta Mall
28mData Curation: Transforming Bytes into AI Gold