Accuracy after forgetting: three seeds, three request sequences
Completed study · 10 October 2026 · One source learning order · Seeds 0, 1, 2
Main finding: the model learns task 15 after forgetting. Its accuracy depends on the seed and request sequence. The forgotten tasks stay at chance accuracy, but some other tasks lose accuracy.
Design
For each seed, we copy the same source model into three branches: learn 15 only (baseline), forget 3 then learn 15, and forget 14 then learn 15. A seed sets random choices, including initialization.
Task 3 was learned first; task 14 was learned last. Compare branches within the same seed.
| Data and model | Learning | Forgetting |
|---|---|---|
| Tiny ImageNet; 10 classes/task; 5,000 train + 500 validation images/task; generated ResNet-50 | 5 epochs; batch 64; Adam, learning rate 0.0001; protection weight 0.1 | 100 steps; 10 noise samples/step; Adam, learning rate 0.0001; noise weight 0.01 |
1. Learning task 15
Accuracy is the percentage of validation images classified correctly. Higher is better.
| Seed | Learn 15 only | Forget 3 → Learn 15 | Forget 14 → Learn 15 |
|---|---|---|---|
| 0 | 50.8% | 50.8% | 48.2% |
| 1 | 44.0% | 47.2% | 47.8% |
| 2 | 34.4% | 47.4% | 32.6% |
| Request sequence | Task-15 changes: seeds 0, 1, 2 | Mean task-15 change | Task-15 SD | Mean other-task change |
|---|---|---|---|---|
| Forget 3 → Learn 15 | +0.0, +3.2, +13.0 | +5.40 | 6.77 | -1.29 |
| Forget 14 → Learn 15 | -2.6, +3.8, -1.8 | -0.20 | 3.49 | -1.02 |
All changes are percentage points: branch minus baseline. Positive means higher accuracy. Means use three seeds. SD (standard deviation) measures variation between seeds; it is not a confidence interval. Other-task means exclude the forgotten task.
Forget 3: the task-15 score matches or exceeds the baseline in every seed. The +5.40-point average includes a large +13.0-point seed-2 effect. Forget 14: gains and losses average to −0.20 points.
2. Forgetting persists; other tasks can suffer
| Task | Before forgetting: seeds 0, 1, 2 | After forgetting | After learning 15 |
|---|---|---|---|
| 3 | 35.6%, 40.4%, 38.4% | 10% in every seed | 10% in every seed |
| 14 | 42.6%, 40.4%, 41.8% | 10% in every seed | 10% in every seed |
Each task has 10 classes, so chance accuracy is 10%. All six comparisons meet the 12% forgetting threshold; no saved evaluation during task-15 learning exceeds it.
| Seed | Forget 3: largest other-task loss | Forget 14: largest other-task loss |
|---|---|---|
| 0 | Task 9: 11.0 points | Task 17: 14.2 points |
| 1 | Task 1: 11.6 points | Task 0: 14.0 points |
| 2 | Task 0: 4.2 points | Task 9: 2.6 points |
Average changes on other tasks are −1.29 points after forgetting task 3 and −1.02 after forgetting task 14. Individual losses reach 14.2 points. A small average does not mean every task is preserved.
Conclusion and limits
The results support forgetting that persists during one later task, with continued ability to learn. They do not show uniform accuracy preservation. Only three seeds, one source order, and one incoming task were tested. Task identity and learning position change together. Chance accuracy does not prove information erasure, privacy, or resistance to recovery.
12/12 jobs · 6/6 comparisons · 39/39 requests complete. Eleven verified session logs total 15 h 10 m 20 s of runner time, including evaluation and transfers; this is not billed GPU time. This diagnostic study uses protection weight 0.1 rather than the repository's paper-based default of 0.01.
Verified scores and settings · Detailed report. All recorded pairing checks pass. Historical results are excluded from these averages.