What is Catastrophic Forgetting
The problem is when a model, after training a new task, sharply loses quality on old tasks.
Definition
Catastrophic Forgetting is a problem when a model, after learning a new task, suddenly loses performance on old tasks. Simply put, this concept helps train models, compare approaches, and reduce the risk of errors on new data. In practice, it helps to understand what capabilities the tool actually has, what data it will need, and what limitations are worth checking before implementation.
Example
The model was additionally trained on new documents, and it began to respond worse to the previous knowledge base.
Why it matters
This is important for systems that must learn gradually and not break already working skills. This helps you choose AI tools not by big promises, but by how they work in a real problem.
How it works
First, the problem is translated into data and metrics, then the model is trained, tested on a separate sample, and compared with alternatives. In the case of the term “Catastrophic Forgetting”, it is important to look separately at the data, quality criteria and application conditions.
Where it is used
- Used in training, testing and tuning models, in automatic selection of parameters, forecasting, classification and recommendation systems.
Limitations
The main limitation is the dependence on data, metrics and verification conditions. A good result on a test does not always mean reliable performance in a real product.