# iCaRL: Incremental Classifier and Representation Learning

**Authors:** Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Christoph H. Lampert  
**arXiv:** [1611.07725](https://arxiv.org/abs/1611.07725)  
**Date:** 2016-11-23  
**Summary page:** https://www.summarizepaper.com/en/paper/1611.07725v1/

## Key points

- Developing systems that can learn incrementally from a continuous stream of data is a challenge in artificial intelligence.
- iCaRL is a novel training strategy that enables class-incremental learning.
- iCaRL learns strong classifiers and data representation simultaneously, making it compatible with deep learning architectures.
- iCaRL outperforms other strategies on two datasets - CIFAR-100 and ImageNet ILSVRC 2012 - by learning many classes incrementally over an extended period without failing quickly.
- iCaRL represents a significant step towards developing intelligent systems capable of continuously expanding their knowledge base from a stream of data.

## Summary

In the quest for artificial intelligence, one of the biggest challenges is developing systems that can learn incrementally from a continuous stream of data and expand their knowledge over time. To address this problem, Sylvestre-Alvise Rebuffi, Alexander Kolesnikov and Christoph H. Lampert introduce iCaRL - a novel training strategy that enables class-incremental learning. Unlike previous approaches which are limited to fixed data representations, iCaRL simultaneously learns strong classifiers and data representation making it compatible with deep learning architectures. The authors demonstrate the effectiveness of iCaRL on two datasets - CIFAR-100 and ImageNet ILSVRC 2012 - where it outperforms other strategies by learning many classes incrementally over an extended period without failing quickly. Overall, iCaRL represents a significant step towards developing intelligent systems capable of continuously expanding their knowledge base from a stream of data.