Curriculum generation using Autoencoder based continuous optimization

Abstract

This paper presents a novel approach to curriculum learning that leverages autoencoder-based continuous optimization for generating effective training sequences. Curriculum learning, which structures the training process from easier to harder examples, has shown promise in improving model performance and convergence. Our method automatically generates curricula by learning a continuous representation of task difficulty through an autoencoder architecture, allowing for smooth progression in training complexity.

Summary

We introduce an automated curriculum generation framework that addresses key challenges in curriculum learning:

• Automatic difficulty assessment of training examples
• Continuous optimization of training sequences
• Smooth progression from simple to complex concepts

The approach uses autoencoders to learn compressed representations of training examples, which are then used to estimate task difficulty and generate optimal training sequences. Our experimental results demonstrate improved learning efficiency and final model performance across multiple tasks and domains.

This work contributes to the growing field of curriculum learning by providing a principled approach to curriculum generation that can be applied to a wide range of machine learning problems.