Machine Learning Methods
Prompt-based method outperforms earlier continual-learning methods on image classification benchmarks
Continual learning systems often forget earlier tasks when they learn new ones, a problem researchers call catastrophic forgetting. In a 2022 paper at the 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), researchers at Northeastern University and Google describe a method called L2P that keeps small learnable instructions known as prompts in a memory space. In image classification benchmarks with different continual learning settings, L2P performed consistently better than earlier methods, and it matched rehearsal-based methods without keeping a rehearsal buffer of old examples.
What problem does continual learning face?
Continual learning trains a model on tasks that arrive one after another, often from changing data distributions. The central challenge is forgetting: adapting model parameters to new data can degrade the model on earlier tasks. Existing methods often rely on a rehearsal buffer, a stored set of earlier examples, or on knowing which task is being tested.
How does L2P work?
L2P treats prompts as small learnable parameters kept in memory. It learns to select prompts dynamically so they instruct a pre-trained model, a model already trained on broad data, to make the right prediction as tasks change. The goal is to manage knowledge that stays useful across tasks and knowledge tied to a particular task while keeping the model able to learn new tasks.
What did the experiments find?
The paper reports tests on popular image classification benchmarks under several challenging continual learning settings. L2P performed consistently better than the earlier methods it was compared with. It also stayed competitive with rehearsal-based methods without using a rehearsal buffer, and the authors report that it applied directly to challenging task-agnostic continual learning, where the model does not see which task it is being asked to solve.
Frequently asked questions
Why is matching rehearsal-based methods notable?
Many earlier continual learning systems need a rehearsal buffer, a stored set of past examples, to hold on to old knowledge. L2P stayed competitive even without that buffer, which matters for settings where storing or replaying earlier data is difficult or unwanted.
What does task-agnostic mean?
In task-agnostic continual learning, the model does not receive a label telling it which task it is performing at test time. The abstract reports that L2P applied directly to that challenging setting.
Limits and what's next
These results come from one paper and one set of experiments. The abstract reports tests on popular image classification benchmarks under different continual learning settings, so the evidence is tied to those benchmark conditions rather than to other kinds of data or real-world use. The abstract reports no independent replication. It does not state how L2P performs outside image classification benchmarks or in other continual learning conditions. The abstract does not say what future work the authors plan.
Source: Zifeng Wang et al., “Learning to Prompt for Continual Learning,” 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022.