Catastrophic forgetting, measured

A small network lost 12.3 percentage points of accuracy on a task it had already mastered, simply by learning a second one. That is real forgetting, and it still missed the registered threshold for what I called severe. A protection method invented in 2017 recovered 98% of the loss without slowing the new task down.

Key takeaways

• Plain sequential fine-tuning dropped Task A accuracy from 97.88% to 85.55%, a 12.3 point loss, below the 20-point bar I registered for "severe" forgetting.
• Elastic Weight Consolidation (EWC) recovered 98.0% of that loss, bringing Task A back to 97.63%, while Task B accuracy stayed within half a point of the unprotected run.
• The registered gate for severe forgetting fired and rejected that hypothesis; the gate for EWC's effectiveness passed by a wide margin. Both verdicts stand as registered, neither was adjusted after seeing the numbers.
• Permuted MNIST, the benchmark used here, shares one output space across both tasks, which likely explains why forgetting was real but moderate rather than catastrophic in the dramatic sense the word usually implies.
• The whole experiment, three training runs on a small network, ran in 26.3 seconds on a consumer GPU. Measuring this phenomenon does not require a large model.

What did the smallest version of this experiment actually measure?

Two tasks, one small network, no tricks. Task A is ordinary MNIST digit classification. Task B is the same images with a fixed random permutation applied to the pixels, a standard continual-learning setup because it keeps the output space identical, ten digit classes, while forcing the network to learn a genuinely different input mapping. The model is a plain 784-400-400 multilayer perceptron, the same scale used in the original Elastic Weight Consolidation paper.

Three training runs from the same starting point after Task A: one continues onto Task B with no protection at all, one continues onto Task B with EWC penalising drift away from the parameters that mattered for Task A. Both are evaluated on the Task A test set afterwards, which is the number that reveals whether anything was forgotten. Full code, five epochs per task, ordinary Adam, is public, so the exact setup behind every number below can be checked rather than taken on trust.

Why did the "severe forgetting" gate not fire?

The registered threshold said a drop of at least 20 percentage points would count as severe forgetting worth calling catastrophic. The measured drop was 12.3 points. The gate fired in the sense that matters here, the sense described in why I pre-register a kill-gate for every experiment: it rejected the hypothesis, cleanly, without argument about where exactly the line should have been drawn, because the line was drawn before the number existed.

The honest explanation is about the benchmark, not the phenomenon. Permuted MNIST is the easy end of continual learning: both tasks share the same ten-class output head, so the network never has to unlearn what a correct answer even looks like, only how to map new inputs onto that unchanged structure. Harder setups, Split MNIST with disjoint output heads for each task being the standard example, tend to produce far more severe collapse, sometimes down to near-chance accuracy on the earlier task. Five epochs with Adam on a small network, in this easier setting, was real but moderate erosion, not the dramatic failure the word "catastrophic" usually conjures. That is a finding about where the severity threshold actually sits for this exact setup, not a reason to have moved the threshold after the fact.

What did EWC actually do differently?

EWC estimates which parameters mattered most for Task A, using a diagonal approximation of the Fisher information computed from a sample of Task A data, then penalises the second training run for moving those specific parameters far from where they started. Here that estimate came from just 1,000 sampled examples, a fraction of the training set, and it still recovered 98.0% of the lost accuracy, landing at 97.63% against the original 97.88%.

The more interesting number sits next to it: Task B accuracy under EWC was 97.43%, against 97.84% for the unprotected run. Protecting Task A cost less than half a point on Task B. The method is not trading one task for the other in any way visible at this scale; it is finding parameter space where both fit. That is close to the original paper's central claim, reproduced here on a network small enough to train in seconds rather than the hours or days a production-scale replication would need.

FAQ

Does a 12.3-point drop not count as catastrophic forgetting at all?

It counts as forgetting. Whether it counts as catastrophic was exactly what the registered gate existed to decide, and by the threshold set in advance, it does not clear that bar. The distinction matters more than it sounds: real, measurable degradation and severe, headline-worthy collapse are different claims, and conflating them is precisely what pre-registration is meant to prevent.

Why use Permuted MNIST instead of a harder continual learning benchmark?

Because the goal was the smallest reproducible demonstration, and Permuted MNIST is the standard entry point the original EWC paper itself used. A harder benchmark, Split MNIST with separate output heads per task, is a natural next registered experiment, and would very likely show more severe forgetting under the same 20-point bar.

Does EWC's near-total recovery mean the forgetting problem is basically solved?

Not on this evidence. This is one small network, one easy benchmark, one hyperparameter setting for EWC's penalty strength, chosen without a search. Whether the same near-total recovery holds for larger models, longer task sequences, or harder benchmarks is a real open question this experiment does not answer.

Where is the full code and data?

In the same public experiment repository as every other registered experiment on this site: the frozen plan, the training script, and the raw metrics file behind every number in this article, so the setup can be checked line by line rather than taken on trust.

References

• James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A. Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, Demis Hassabis, Claudia Clopath, Dharshan Kumaran, Raia Hadsell. Overcoming catastrophic forgetting in neural networks, 2016. arxiv.org/abs/1612.00796