<core_client_version>8.2.8</core_client_version> <![CDATA[ <stderr_txt> 2026-03-14 07:35:18 (10180): Streaming wrapper: running ../../projects/axiom.heliex.net/axiom_experiment_634_cpu.exe (wu.json result.json) 2026-03-14 07:35:18 (10180): Streaming wrapper: created child process 4168 [CPU] Using NumPy [BOINC] Starting BOINC mode [BOINC] Work unit: wu.json [BOINC] Result file: result.json [BOINC] Current directory: C:\ProgramData\BOINC\slots\1 [BOINC] Directory contents: ['.cupy_cache', 'axiom_experiment_634_cpu.exe', 'boinc_lockfile', 'boinc_setup_complete', 'init_data.xml', 'job.xml', 'result.json', 'stderr.txt', 'streaming_wrapper_259_windows_x86_64.exe', 'streaming_wrapper_checkpoint.txt', 'wu.json'] [BOINC] Resolved soft link: ../../projects/axiom.heliex.net/exp_power_law_forgetting_v2_d1d3d4.json [BOINC] Resolved soft link: ../../projects/axiom.heliex.net/exp_power_law_forgetting_v2_d1d3d4_0_r1802744365_0 [BOINC] Actual WU file: ../../projects/axiom.heliex.net/exp_power_law_forgetting_v2_d1d3d4.json [BOINC] Successfully loaded work unit [BOINC] Experiment container mode detected [EXP] Experiment mode: exp_power_law_forgetting_v2_d1d3d4 [EXP] Script URL: https://axiom.heliex.net/experiments/power_law_forgetting_v2.py [EXP] Run duration: 900s [EXP] Downloading script... [EXP] Downloaded 20274 bytes [EXP] Executing experiment 'exp_power_law_forgetting_v2_d1d3d4'... [EXP] No GPU available, running CPU-only ====================================================================== POWER-LAW FORGETTING v2 Catastrophic Forgetting with Overlapping Representations ====================================================================== Input dim: 20 Output classes: 10 Hidden layer 1: 50 Bottleneck widths: [20, 50, 100] Training samples per task: 1000 Test samples per task: 300 Task B epochs: 300 EWC lambda: 1000.0 Task mapping cosine similarity: -0.0709 (0 = orthogonal, 1 = identical) --- bottleneck=20_naive --- Architecture: 20 -> 50 -> 20 -> 10 (2280 params) Phase 1: Training Task A ... done (500 epochs, 1.0s) Task A accuracy: 0.7200 Task B accuracy: 0.0933 Phase 2: Training Task B, measuring forgetting ... Done (0.8s) Final Task A: 0.0767 (forgot 0.6433) Final Task B: 0.7867 --- bottleneck=20_ewc --- Architecture: 20 -> 50 -> 20 -> 10 (2280 params) Phase 1: Training Task A ... done (500 epochs, 1.7s) Task A accuracy: 0.7200 Task B accuracy: 0.0933 Computing Fisher Information ... done (0.1s) Phase 2: Training Task B, measuring forgetting ... Done (1.1s) Final Task A: 0.3933 (forgot 0.3267) Final Task B: 0.1933 --- bottleneck=50_naive --- Architecture: 20 -> 50 -> 50 -> 10 (4110 params) Phase 1: Training Task A ... done (500 epochs, 2.2s) Task A accuracy: 0.7333 Task B accuracy: 0.0733 Phase 2: Training Task B, measuring forgetting ... Done (1.2s) Final Task A: 0.0700 (forgot 0.6633) Final Task B: 0.7700 --- bottleneck=50_ewc --- Architecture: 20 -> 50 -> 50 -> 10 (4110 params) Phase 1: Training Task A ... done (500 epochs, 2.0s) Task A accuracy: 0.7333 Task B accuracy: 0.0733 Computing Fisher Information ... done (0.1s) Phase 2: Training Task B, measuring forgetting ... Done (1.4s) Final Task A: 0.2633 (forgot 0.4700) Final Task B: 0.4133 --- bottleneck=100_naive --- Architecture: 20 -> 50 -> 100 -> 10 (7160 params) Phase 1: Training Task A ... done (500 epochs, 2.8s) Task A accuracy: 0.7333 Task B accuracy: 0.0767 Phase 2: Training Task B, measuring forgetting ... Done (1.5s) Final Task A: 0.0800 (forgot 0.6533) Final Task B: 0.8033 --- bottleneck=100_ewc --- Architecture: 20 -> 50 -> 100 -> 10 (7160 params) Phase 1: Training Task A ... done (500 epochs, 2.6s) Task A accuracy: 0.7333 Task B accuracy: 0.0767 Computing Fisher Information ... done (0.1s) Phase 2: Training Task B, measuring forgetting ... Done (1.8s) Final Task A: 0.1433 (forgot 0.5900) Final Task B: 0.5333 ====================================================================== RESULTS SUMMARY ====================================================================== Total runtime: 20.4s Condition AccA_init AccA_final Forgetting Power_exp ---------------------------------------------------------------------- bottleneck=20_naive 0.7200 0.0767 0.6433 0.477 bottleneck=20_ewc 0.7200 0.3933 0.3267 0.524 bottleneck=50_naive 0.7333 0.0700 0.6633 0.483 bottleneck=50_ewc 0.7333 0.2633 0.4700 0.590 bottleneck=100_naive 0.7333 0.0800 0.6533 0.418 bottleneck=100_ewc 0.7333 0.1433 0.5900 0.482 Key findings: Bottleneck=20: EWC reduces forgetting by 0.3167 (0.6433 -> 0.3267) Bottleneck=50: EWC reduces forgetting by 0.1933 (0.6633 -> 0.4700) Bottleneck=100: EWC reduces forgetting by 0.0633 (0.6533 -> 0.5900) Capacity effect (naive SGD): Width 20: forgetting = 0.6433 Width 50: forgetting = 0.6633 Width 100: forgetting = 0.6533 Results saved to experiment_result.json [EXP] Read experiment_result.json successfully [EXP] Experiment 'exp_power_law_forgetting_v2_d1d3d4' finished in 20.5s [EXP] Status: completed 2026-03-14 07:35:41 (10180): axiom_experiment_634_cpu.exe exited; CPU time 18.765625 2026-03-14 07:35:41 (10180): called boinc_finish(0) </stderr_txt> ]]>