Rocket Launching: A Universal and Efficient Framework for Training Well-performing Light Net

Abstract

Models applied on real time response task, like click-through rate (CTR) prediction model, require high accuracy and rigorous response time. Therefore, top-performing deep models of high depth and complexity are not well suited for these applications with the limitations on the inference time. In order to further improve the neural networks' performance given the time and computational limitations, we propose an approach that exploits a cumbersome net to help train the lightweight net for prediction. We dub the whole process rocket launching, where the cumbersome booster net is used to guide the learning of the target light net throughout the whole training process. We analyze different loss functions aiming at pushing the light net to behave similarly to the booster net, and adopt the loss with best performance in our experiments. We use one technique called gradient block to improve the performance of the light net and booster net further. Experiments on benchmark datasets and real-life industrial advertisement data present that our light model can get performance only previously achievable with more complex models.

arXiv comment10 pages, AAAI2018

Cited by in corpus (18)

• Knowledge Distillation: A Survey
• Low-resolution Face Recognition in the Wild via Selective Knowledge Distillation
• Knowledge Distillation in Deep Learning and its Applications
• $F{M}^2$: Field-matrixed Factorization Machines for Recommender Systems
• Knowledge Distillation via Route Constrained Optimization
• DBP: Discrimination Based Block-Level Pruning for Deep Model Acceleration
• Knowledge distillation via adaptive instance normalization
• Knowledge Squeezed Adversarial Network Compression
• User Response Prediction in Online Advertising
• MC${}^2$-SF: Slow-Fast Learning for Mobile-Cloud Collaborative Recommendation
• DADNN: Multi-Scene CTR Prediction via Domain-Aware Deep Neural Network
• Triplet Distillation for Deep Face Recognition
• Progressive Learning of Low-Precision Networks
• Knowledge Representing: Efficient, Sparse Representation of Prior Knowledge for Knowledge Distillation
• A Hierarchical User Intention-Habit Extract Network for Credit Loan Overdue Risk Detection
• Privileged Features Distillation at Taobao Recommendations
• Learning Compatible Embeddings
• Fine-grained Knowledge Fusion for Sequence Labeling Domain Adaptation