Overview
Learn essential techniques for debugging neural networks in natural language processing applications. This comprehensive lecture covers identifying problems, addressing training time issues, and resolving test time challenges. Explore strategies for optimizing model size, implementing residual connections, fine-tuning optimizers and learning rates, and improving initialization. Discover methods for effective mini-batching, learning rate decay, and battling overfitting. Gain insights into debugging test time performance, including minibatch bugs, unit testing, beam search, and output generation. Master quantitative analysis techniques and compare empty toolkit approaches to enhance your neural network debugging skills for NLP tasks.
Syllabus
Introduction
Problems with Neural Networks
Key to Debugging
Problem
Possible Causes
Debugging Training Time
Model Size
Residual Connections Highway Networks
Optimization
Optimizers
Learning Rate
Initialization
Mini Batching
Sorting
Learning Rate Decay
Other Questions
Test Time Performance
Minibatch Bugs
Unit Testing
Beam Search
Output Generation
Quantitative Analysis
Compare Empty Toolkit
Battling Overfitting
Memory
Taught by
Graham Neubig