Curriculum Learning

Introduction Curriculum learning accelerates the convergence of a model, and helps to find a better local minimum for non-convex problems. Here we summarize a few works. Curriculum Learning Problem Introduce curriculum learning. Formalize such training... Continue

NLP Benchmarks

DecaNLP The Natural Language Decathlon is a multitask challenge that spans 10 tasks. In their model, they treat everything as question answering, which seems very nice to me. The following table shows the tasks included... Continue

Grow a Neural Network

Introduction Automatically evolving a network structure for handling different tasks is an important research problem. Here we summarize a few works. Learn to Grow: A Continual Structure Learning Framework for Overcoming Catastrophic Forgetting Problem Handling... Continue

Paper Reading - Multi-task Learning over Graph Structures

Problem How to learn complex relationships between different tasks. Key Ideas Instead of use flat structure or pre-defined hierarchical structure, learn a complete graph or star graph between tasks. It is weighted directed graph. Tasks... Continue