Pinterest engineers enable multi-epoch embedding training

Pinterest engineers detail a method to enable safe and efficient multi-epoch embedding training for large ID embedding tables in their Homefeed ranking model. By keeping embedding tables frozen initially while allowing the rest of the network to train, the team avoided catastrophic overfitting and achieved a 1.09% gain in online saves. To make extended training practical, they also introduced a semi-synchronous training pipeline and optimized lookup operations to reduce GPU idle time. Open article →