Does Learning Order Explain Forgetting? The Lie Bracket as an Early-Warning Signal for Catastrophic Forgetting Without Old Data
Organizations: Responsible Artificial Intelligence Lab (RAIL), Kwame Nkrumah University of Science and Technology (KNUST), Ghana
Abstract
A network trained on task A and then task B ends at different weights than one trained on B and then A. To leading order this gap is set by the Lie bracket of the two tasks' gradient fields, and Dherin (2023) conjectured that a non-vanishing bracket may be linked to catastrophic forgetting. I test this conjecture empirically. Across Split MNIST, Split Fashion-MNIST and Split CIFAR-10, the forgetting of A is almost exactly the difference in A's loss between the two training orders (Spearman rho = 0.96-1.00). I then ask whether the bracket can predict that forgetting early in training on B and without any of A's data, using only a small summary stored when A finishes: its gradient and its top 20 Hessian eigenpairs. When the old task is well learned and forgetting is slow (task-incremental), this stored bracket ranks forgetting with rho = 0.70-0.81, against 0.33-0.58 for the penalty of Elastic Weight Consolidation, and matches or exceeds re-evaluating A on its own data. When forgetting is fast (domain-incremental), the stored summary goes stale and the EWC penalty is as good or better. The bracket fails outright when the old task is only half learned; I trace this to a residual gradient on A an order of magnitude larger, which makes first-order (slope) damage matter. The bracket is therefore a useful early-warning signal for the quiet kind of forgetting, under a clear condition.
Figures & tables
| Setup | Input | Model | Accuracy | Forgetting, % mean (max) | |
|---|---|---|---|---|---|
| slow | fast | ||||
| MNIST | pixels (784) | MLP 256–256 | 0.989 | 1.3 (7.1) | 35.0 (85.2) |
| Fashion-MNIST | pixels (784) | MLP 256–256 | 0.990 | 0.9 (8.9) | 19.8 (46.5) |
| CIFAR-10 features | frozen ResNet-18 (512) | MLP 256–256 | 0.973 | 0.7 (3.5) | 18.6 (46.2) |
| CIFAR-10 CNN | pixels ( ) | small CNN | 0.776 | 4.1 (14.1) | 21.2 (43.4) |
| MNIST | Fashion-MNIST | CIFAR-10 features | CIFAR-10 CNN (half learned) | |
|---|---|---|---|---|
| Forgetting vs. order gap on | 1.00 / 1.00 | 0.99 / 1.00 | 0.99 / 1.00 | 0.96 / 0.99 |
| Order gap forgetting | 1.04 / 1.00 | 1.00 / 1.00 | 1.00 / 1.00 | 1.20 / 1.05 |
| Bracket vs. order gap on | 0.83 / 0.69 | 0.68 / 0.69 | 0.79 / 0.81 | 0.03 / 0.29 |
| Bracket vs. weight gap | 0.66 / 0.69 | 0.06 / 0.58 | 0.54 / 0.48 | 0.16 / 0.39 |
| MNIST | Fashion-MNIST | CIFAR-10 features | CIFAR-10 CNN (half learned) | |
| Slow forgetting (Task-IL) | ||||
| Stored bracket (proposed) | 0.81 [0.70, 0.88] | 0.70 [0.56, 0.81] | 0.76 [0.64, 0.84] | 0.12 [ 0.31, 0.08] |
| EWC penalty | 0.58 [0.42, 0.71] | 0.33 [0.16, 0.48] | 0.35 [0.13, 0.55] | 0.22 [0.03, 0.39] |
| Stored Taylor estimate | 0.67 [0.52, 0.79] | 0.31 [0.10, 0.51] | 0.58 [0.40, 0.72] | 0.32 [0.12, 0.49] |
| Full bracket † | 0.82 [0.72, 0.88] | 0.74 [0.61, 0.84] | 0.76 [0.65, 0.84] | 0.11 [ 0.30, 0.09] |
| Re-evaluation † | 0.81 [0.71, 0.87] | 0.33 [0.11, 0.53] | 0.68 [0.53, 0.79] | 0.31 [0.11, 0.49] |
| MNIST | Fashion-MNIST | CIFAR-10 features | CIFAR-10 CNN (half learned) | |
|---|---|---|---|---|
| at | 0.07 / 0.03 | 0.04 / 0.02 | 0.07 / 0.04 | 0.94 / 0.38 |
| Slope share of damage | 0.31 / 0.11 | 0.35 / 0.12 | 0.25 / 0.10 | 0.57 / 0.23 |
| Full bracket | 0.82 / 0.69 | 0.67 / 0.69 | 0.78 / 0.80 | 0.04 / 0.30 |
| term alone | 0.78 / 0.79 | 0.71 / 0.77 | 0.79 / 0.75 | 0.02 / 0.19 |
| term alone | 0.75 / 0.68 | 0.50 / 0.63 | 0.68 / 0.72 | 0.04 / 0.25 |
| Measured curvature damage | 0.84 / 0.47 | 0.71 / 0.80 | 0.87 / 0.78 | 0.17 / 0.73 |
Appendix figures & tables6 assets
Supplementary material from the paper’s appendix.
Appendix
| Fashion-MNIST (MLP) | 0.878 (0.951) | 0.979 (0.995) | 0.995 (0.999) | 0.999 (1.000) |
|---|---|---|---|---|
| CIFAR-10 features (MLP) | 0.948 (0.996) | 0.984 (1.000) | 0.995 (1.000) | 0.999 (1.000) |
| CIFAR-10 CNN | 1.005 (0.933) | 0.992 (0.994) | 0.996 (0.999) | 0.999 (1.000) |
| Stored bracket | EWC | Taylor | Full bracket † | Re-eval. † | Forgetting | ||
|---|---|---|---|---|---|---|---|
| Slow | 0.02 | 0.86 | 0.55 | 0.58 | 0.86 | 0.61 | 1.0 |
| Slow | 0.05 | 0.81 | 0.58 | 0.67 | 0.82 | 0.81 | 1.3 |
| Slow | 0.1 | 0.71 | 0.59 | 0.79 | 0.73 | 0.84 | 1.2 |
| Fast | 0.02 | 0.46 | 0.61 | 0.46 | 0.72 | 0.77 | 30.9 |
| Fast | 0.05 | 0.40 | 0.59 | 0.41 | 0.69 | 0.81 | 35.0 |
| Fast | 0.1 | 0.33 | 0.49 | 0.37 | 0.63 | 0.82 | 37.1 |
| EWC | Stored br. | Br 10 | Chk 10 | Br 50 | Chk 50 | Br 200 | Chk 200 | ||
|---|---|---|---|---|---|---|---|---|---|
| Slow | 0.02 | 0.55 | 0.86 | 0.73 | 0.01 | 0.79 | 0.14 | 0.83 | 0.24 |
| Slow | 0.05 | 0.58 | 0.81 | 0.66 | 0.12 | 0.71 | 0.31 | 0.77 | 0.42 |
| Slow | 0.1 | 0.59 | 0.71 | 0.57 | 0.11 | 0.65 | 0.32 | 0.69 | 0.59 |
| Fast | 0.02 | 0.61 | 0.46 | 0.74 | 0.68 | 0.75 | 0.71 | 0.77 | 0.76 |
| Fast | 0.05 | 0.59 | 0.40 | 0.72 | 0.71 | 0.72 | 0.79 | 0.72 | 0.81 |
| Fast | 0.1 | 0.49 | 0.33 | 0.57 | 0.72 | 0.57 | 0.79 | 0.59 | 0.82 |
| Stored br. | Br 10 | Br 50 | Chk 10 | Chk 50 | EWC | ||
|---|---|---|---|---|---|---|---|
| Fashion-MNIST | slow | 0.70 | 0.41 | 0.52 | 0.29 | 0.29 | 0.33 |
| Fashion-MNIST | fast | 0.78 | 0.59 | 0.66 | 0.59 | 0.69 | 0.79 |
| CIFAR-10 features | slow | 0.76 | 0.29 | 0.62 | 0.35 | 0.06 | 0.35 |
| CIFAR-10 features | fast | 0.46 | 0.28 | 0.69 | 0.22 | 0.65 | 0.84 |
| CIFAR-10 CNN | slow | 0.12 | 0.12 | 0.21 | 0.11 | 0.19 | 0.22 |
| CIFAR-10 CNN | fast | 0.30 | 0.12 | 0.28 | 0.48 | 0.64 | 0.47 |
| Signal (no old data unless stated) | with forgetting |
|---|---|
| EWC penalty | 0.84 |
| Stored bracket, | 0.46 |
| Stored bracket EWC penalty | 0.71 |
| Stored bracket Taylor estimate | 0.62 |
| Stored bracket EWC penalty Taylor estimate | 0.75 |
| Stored bracket, | 0.49 |
| MNIST | Fashion-MNIST | CIFAR-10 features | CIFAR-10 CNN (half learned) | |
| Largest eigenvalue | 5.43 | 3.85 | 2.35 | 16.26 |
| Smallest eigenvalue | ||||
| 0.065 | 0.072 | 0.051 | 0.150 | |
| 0.59 | 0.62 | 0.59 | 0.82 |
Catastrophic Forgetting
Catastrophic forgetting is the loss of previously learned behavior when a model adapts to new tasks or changing data.
See all 374 papers