Specialty Digest

Global English editionLearners’ Study Dashboard — in development
The Knowledge Library for All

Continual Learning

nounMachine Learningalso lifelong learning, incremental learning, sequential learning

In one line

Continual learning is a machine learning approach in which a model learns from a sequence of tasks or data over time while retaining what it learned earlier.

Continual learning is a machine learning approach in which a model learns from a stream of tasks or data over time, adding new knowledge while keeping what it learned before. People do this naturally. A person who learns to ride a bicycle does not forget how to walk. Neural networks, by contrast, tend to lose old skills when trained on new material, a failure known as catastrophic forgetting. Continual learning, also called lifelong or incremental learning, tries to overcome this problem so that AI systems can keep improving after deployment without being retrained from scratch on all their data. It is a close relative of online learning and a major goal of deep learning research.

Why continual learning matters

The world changes, and so do the tasks a system faces. A home robot meets new rooms and objects, a language model faces new facts and a medical system sees new diseases and devices. The standard solution is to gather all old and new data together and retrain, which is expensive in computing and storage and sometimes impossible when old data cannot be kept because of privacy rules or limited memory. Continual learning promises systems that update efficiently, adapt to individual users and accumulate knowledge across their working life.

The problem of catastrophic forgetting

In a neural network, knowledge is stored in the same set of weights for all tasks. When the network trains on a new task, gradient descent changes those weights to fit the new data, and the changes can destroy the settings that supported earlier tasks. The effect was described by Michael McCloskey and Neal Cohen in 1989 and by Roger Ratcliff in 1990. It is a form of the stability-plasticity dilemma. A model needs plasticity to learn new things and stability to keep old ones, and too much of either causes trouble. Continual learning is the search for a good balance between the two.

Main approaches

  • Replay methods: the system keeps a small memory of old examples, or generates pseudo-examples with a generative model, and mixes them into training on new tasks so that old skills stay fresh. Experience replay is simple and often the strongest baseline.
  • Regularization methods: these add a penalty that discourages changes to weights that mattered for earlier tasks. Elastic weight consolidation, proposed by DeepMind researchers in 2017, estimates how important each weight is and protects the important ones. Learning without forgetting keeps the model’s outputs on old tasks close to their earlier values.
  • Architecture methods: the network grows or is partitioned so that each task gets its own parameters. Progressive neural networks add new columns for new tasks, and methods such as PackNet assign different subsets of weights to different tasks.
  • Parameter-efficient methods: small adapter modules or prompts are trained per task on top of a frozen pretrained model, which avoids overwriting the base knowledge.
  • Meta-learning approaches: the model is trained to learn in a way that interferes less with earlier knowledge. See meta-learning.

Scenarios and evaluation

Researchers distinguish three common settings. In task-incremental learning, the task identity is given at test time, so the model only chooses among classes within that task. In domain-incremental learning, the task is the same but the input distribution changes. In class-incremental learning, new classes keep arriving and the model must choose among all classes seen so far without being told the task, which is the hardest case. Common measures include average accuracy over all tasks, backward transfer, which shows how much learning new tasks changed performance on old ones, and forward transfer, which shows whether earlier learning helped with later tasks. Benchmarks such as split MNIST, permuted MNIST and split CIFAR are used to compare methods.

Applications of continual learning

Robots and autonomous vehicles need to adapt to new environments without forgetting earlier skills. Personal assistants and recommendation systems adjust to each user’s changing habits. Medical models update with new imaging devices and diseases. Fraud and cybersecurity systems must keep up with evolving attacks. Large language models face the issue of keeping knowledge current, and researchers explore continual pretraining and model editing to add new facts or skills without damaging old ones. Edge devices with limited storage benefit because they cannot keep all past data.

Advantages of continual learning

The approach saves computing and storage compared with full retraining, enables adaptation over time, supports privacy by reducing the need to keep old data and moves AI toward the kind of cumulative learning seen in people and animals.

Challenges and open problems

No method fully solves forgetting, and approaches that work on small benchmarks often struggle on realistic, large-scale problems. Replay needs stored data, which may raise privacy and memory limits. Regularization can limit the capacity to learn new things as constraints accumulate. Architecture growth increases model size. Evaluation is also hard because results depend on the order and similarity of tasks. Real data streams are messy, with no clear task boundaries, and research is now turning to task-free continual learning, to large pretrained models and to the link between continual learning and in-context learning.

A worked example

A company trains an image classifier to sort products into fifty categories. Each quarter, ten new categories appear. Retraining on the entire catalog takes days. Instead, the team keeps a small memory of a few images per old category, trains on the new images mixed with this memory and adds a penalty that keeps important weights from drifting. After each update, they test on all categories seen so far and compare with a model fully retrained on everything. If accuracy on old categories falls by more than a set margin, they enlarge the memory or run a full retraining.

FAQs

What is the difference between continual learning and transfer learning? Transfer learning reuses knowledge from a source task for a target task and cares mainly about the target. Continual learning cares about performing well on all tasks, old and new.

Can large language models learn continually? Not easily. They are usually updated in separate training runs, and research is exploring safer ways to add knowledge without forgetting.

What is a replay buffer? It is a small store of past examples that is mixed into training to remind the model of earlier tasks.

Key facts

  • Continual learning aims to learn from a stream of tasks without retraining on all earlier data.1
  • The main obstacle is catastrophic forgetting, in which training on new data overwrites knowledge of earlier tasks.2
  • Main strategies include replay, regularization of important weights and architecture expansion.1

Related concepts

In the news

Quick checkWhat is catastrophic forgetting?Show answer

The tendency of a neural network to lose performance on earlier tasks when it is trained on new ones.

Sources

  1. De Lange et al., A Continual Learning Survey, https://arxiv.org/abs/1909.08383
  2. Kirkpatrick et al., Overcoming catastrophic forgetting in neural networks, https://arxiv.org/abs/1612.00796
  3. Parisi et al., Continual Lifelong Learning with Neural Networks: A Review, https://arxiv.org/abs/1802.07569

Editorially reviewed by Specialty Digest Editorial TeamLast reviewed November 17, 2026Researched and drafted with AI assistanceReport an issue