# Continual Lifelong Learning with Neural Networks: A Review

**Authors:** German I. Parisi, Ronald Kemker, Jose L. Part, Christopher Kanan, Stefan Wermter  
**arXiv:** [1802.07569](https://arxiv.org/abs/1802.07569)  
**Date:** 2018-02-21  
**Summary page:** https://www.summarizepaper.com/en/paper/1802.07569v4/

## Key points

- Lifelong learning is the ability to continually acquire and transfer knowledge and skills throughout an agent's lifespan.
- Machine learning and neural network models struggle with lifelong learning due to catastrophic forgetting or interference when faced with non-stationary data distributions.
- Various approaches have been proposed to address this limitation, such as regularization techniques, episodic memory-based methods, dynamic architectures, generative replay-based methods among others.
- Biological systems provide inspiration for developing more effective lifelong learning algorithms.
- Structural plasticity allows neurons to form new connections or strengthen existing ones based on experience while memory replay enables consolidation of memories during sleep.
- Curriculum and transfer learning leverages prior knowledge to facilitate faster acquisition of new skills while intrinsic motivation encourages exploration of novel environments.
- Multisensory integration allows different sensory modalities to interact synergistically.

## Summary

This review article focuses on the challenge of lifelong learning for artificial learning systems, which refers to the ability to continually acquire and transfer knowledge and skills throughout an agent's lifespan. While humans and animals possess this ability through a rich set of neurocognitive mechanisms, machine learning and neural network models struggle with it due to catastrophic forgetting or interference when faced with non-stationary data distributions. The authors critically summarize the main challenges linked to lifelong learning for artificial learning systems and compare existing neural network approaches that alleviate catastrophic forgetting. They discuss well-established and emerging research motivated by lifelong learning factors in biological systems such as structural plasticity, memory replay, curriculum and transfer learning, intrinsic motivation, and multisensory integration. The authors highlight the importance of lifelong learning capabilities for autonomous agents interacting in the real world and processing continuous streams of information. However, current deep neural network models typically learn representations from stationary batches of training data without accounting for situations in which information becomes incrementally available over time. To address this limitation, various approaches have been proposed such as regularization techniques, episodic memory-based methods, dynamic architectures, generative replay-based methods among others. The review also discusses how biological systems provide inspiration for developing more effective lifelong learning algorithms. For example, structural plasticity allows neurons to form new connections or strengthen existing ones based on experience while memory replay enables consolidation of memories during sleep. Curriculum and transfer learning leverages prior knowledge to facilitate faster acquisition of new skills while intrinsic motivation encourages exploration of novel environments. Multisensory integration allows different sensory modalities to interact synergistically. Overall, this review provides a comprehensive overview of the challenges associated with achieving lifelong learning in artificial intelligence systems along with potential solutions inspired by biological mechanisms that can be used to develop more effective algorithms for autonomous agents interacting in real-world environments.