Personalization and Scalable Deep Learning with
summary
Claims that time-varying personalization yields better results when modeled with an LSTM-based recurrent latent-variable model that predicts user return times via hazard/survival outputs and captures movie rating spikes around events. Implementation uses a mix of imperative and declarative code on MXNet with an MXNet CloudFormation template to spin up clusters and a parameter-server sync; benchmarks show linear scaling up to 108x speed-up on 128 GPUs (70x for AlexNet) and live runs of hundreds to thousands of samples per GPU, beating extensive manual feature engineering.
AI-written summary · about
chapters (8)
- 0:00Introduction to the session
- 1:37Personalization techniques
- 12:02Survival analysis for engagement
- 15:20Recommender system challenges
- 21:51Deep learning with MXNet
- 27:32Scalable model training
- 31:14Live infrastructure demo
- 37:49Closing and Q&A
related talks
42mAlex Smola, Professor in the Machine Learning Department, Carnegie Mellon University
03mMachine Learning for Personalization
29mFurther enhancements of recommender systems using deep learning
23mBest Practices for Building Production-Grade Deep Learning Systems
18mBuilding an Incrementally Trained, Local Taste Aware, Global Deep Learned Recommender System Model
12mQuickUpdate: a Real-Time Personalization System for Large-Scale Recommendation Models
19mAdvancing Personalized Learning Analytics with AI and EdTech
31mThe Advantages of Deep Learning with Recurrent Neural Networks