learnability dict | continual_pressure dict | bounded_replay_value dict | replay_scaling dict | residual_headroom dict |
|---|---|---|---|---|
{
"joint_every_seed_min": 80,
"joint_cohort_mean_min": 82,
"pass_every_seed": false,
"pass_cohort_mean": false
} | {
"joint_minus_sgd_every_seed_min": 55,
"joint_minus_sgd_cohort_mean_min": 60,
"pass_every_seed": true,
"pass_cohort_mean": true
} | {
"er200_minus_sgd_every_seed_min": 15,
"er200_minus_sgd_cohort_mean_min": 20,
"pass_every_seed": true,
"pass_cohort_mean": true
} | {
"er500_minus_er200_every_seed_min": 5,
"er500_minus_er200_cohort_mean_min": 8,
"er1000_minus_er500_every_seed_min": 5,
"er1000_minus_er500_cohort_mean_min": 8,
"pass_er500_every_seed": true,
"pass_er500_mean": true,
"pass_er1000_every_seed": true,
"pass_er1000_mean": true
} | {
"joint_minus_er1000_every_seed_min": 4,
"joint_minus_er1000_cohort_mean_min": 7,
"pass_every_seed": false,
"pass_cohort_mean": false
} |
MNIST-360 Continual-Learning Qualification
A reproducible qualification study of the actual MNIST-360 General Continual Learning stream using a pinned official Mammoth implementation.
The study prospectively compared five fixed conditions across 12 fresh seeds:
| Condition | Mean final accuracy |
|---|---|
| SGD | 20.37% |
| ER-200 | 47.91% |
| ER-500 | 61.18% |
| ER-1000 | 75.17% |
| Joint/offline | 81.67% |
Result
Negative qualification: NOT_SUFFICIENTLY_LEARNABLE_UNDER_FROZEN_GATES
The first binding gate required joint/offline training to reach at least 80% on every seed and 82% cohort mean. The observed joint mean was 81.67%, with two seeds below 80%.
Other pressures were clearly present: ordinary online SGD was poor, bounded replay added substantial value, and larger replay buffers improved performance. But the substrate failed the prospectively frozen learnability criterion, so the negative result was preserved rather than repaired after inspection.
Reproduce
Open notebooks/RUN_IN_COLAB.ipynb in Google Colab. The notebook clones Mammoth at commit e75a491c69fd729edeb01431afb753d9157d9a81, verifies pinned Git objects, applies and hashes the frozen compatibility patch, runs 60 scientific runs, and evaluates the already-frozen gates.
The public notebook changes only release naming and output paths from the source research artifact. Scientific settings, seeds, thresholds, commits, and run configurations are unchanged.
Repository contents
protocol/FROZEN_PROTOCOL_PUBLIC_RELEASE.md— prospective frozen protocol, public namingnotebooks/RUN_IN_COLAB.ipynb— reproducible Colab runnerresults/raw_results.csv— all 60 observed final accuraciesresults/metrics_by_seed.csv— per-seed derived metricsresults/cohort_summary.csv— cohort descriptive statisticsresults/gate_evaluation.json— frozen gate evaluationresults/qualification_result.md— concise result and claim boundarydocs/SOURCE_CUSTODY.md— source hashes and release modifications
Scope
This artifact evaluates one pinned MNIST-360 substrate and configuration matrix. It does not establish general superiority or inferiority of replay, joint training, SGD, or continual-learning methods.
- Downloads last month
- 41