Revisiting Distillation and Incremental Classifier Learning

TL;DR

This paper critically analyzes the iCaRL method for incremental learning, identifies knowledge distillation as its main success factor, highlights its bias issues, and proposes a dynamic threshold moving algorithm to improve performance on CIFAR100 and MNIST.

Contribution

It isolates the core effective idea of knowledge distillation in incremental learning and introduces a novel bias correction algorithm to enhance its effectiveness.

Findings

iCaRL's success is mainly due to knowledge distillation

Knowledge distillation can introduce classifier bias

The proposed dynamic threshold moving algorithm reduces bias and improves results

Abstract

One of the key differences between the learning mechanism of humans and Artificial Neural Networks (ANNs) is the ability of humans to learn one task at a time. ANNs, on the other hand, can only learn multiple tasks simultaneously. Any attempts at learning new tasks incrementally cause them to completely forget about previous tasks. This lack of ability to learn incrementally, called Catastrophic Forgetting, is considered a major hurdle in building a true AI system. In this paper, our goal is to isolate the truly effective existing ideas for incremental learning from those that only work under certain conditions. To this end, we first thoroughly analyze the current state of the art (iCaRL) method for incremental learning and demonstrate that the good performance of the system is not because of the reasons presented in the existing literature. We conclude that the success of iCaRL is primarily due to knowledge distillation and recognize a key limitation of knowledge distillation, i.e, it often leads to bias in classifiers. Finally, we propose a dynamic threshold moving algorithm that is able to successfully remove this bias. We demonstrate the effectiveness of our algorithm on CIFAR100 and MNIST datasets showing near-optimal results. Our implementation is available at https://github.com/Khurramjaved96/incremental-learning.