Learning Trajectories Leave Decodable Traces in Neural Network Weight Space and Internal Representations
Organizations: Cognitive Computational Neuroscience Group, Pattern Recognition Lab, Friedrich-Alexander-University Erlangen-Nürnberg (FAU), Erlangen, Germany · Neural Interface Lab, University Hospital Mannheim, University Heidelberg, Mannheim, Germany · Neuroscience Lab, University Hospital Erlangen, Erlangen, Germany · BG Clinik Ludwigshafen, Ludwigshafen am Rhein, Germany
Abstract
Neural networks trained on the same task develop related weight-space structures, yet remain distinguishable after learning. Do these differences merely reflect alternative solutions, or do they preserve information about individual learning histories? We tested whether final weights reveal initialization, task order, and earlier trajectory states. Training reorganized the ensemble: initial pairwise distances did not consistently predict final distances. Nevertheless, all 30 trained networks were correctly matched to their initial weights among ten candidates per dataset. Networks sharing an initialization also developed recognizable final hidden representations across different batch schedules, although direct comparisons with initial activity did not recover identity. Learning order remained readable as well. A linear probe decoded the order of two earlier tasks with 100% accuracy on 80 held-out networks after ten epochs of a common final task. Earlier intermediate weight states were partly predictable, but individual initial weights could not be reliably reconstructed. Final weights thus retain selective historical information rather than a complete trajectory record. Perturbed networks recovered classification performance without returning to their unperturbed training paths. This separation persisted across datasets, intervention times, and network layers. Moreover, 2,157 of 2,160 aligned endpoint pairs showed no positive loss barrier at sampled interpolation points, indicating low-loss connections between distinct solutions rather than convergence to one weight vector. Together, these findings show that a shared task constrains networks to related successful solutions without erasing their individual histories. Final weights reflect not only what a network has learned, but also where learning began and how it unfolded.
Figures & tables
| Cohort | Design | Main comparison |
|---|---|---|
| Main cohort | 3 datasets; 10 runs each | Initial–final geometry, weight identity, and representation identity |
| Crossed training | 10 initializations 5 batch schedules per dataset; 150 runs | Shared initialization across schedules; shared schedule across initializations |
| Perturbation pilot | 10 MNIST sources; 300 perturbed continuations plus controls | Disturbed versus unperturbed continuation |
| Perturbation extension | 3 datasets 3 times; 2,160 perturbations and 180 baselines | Generalization across intervention time, layer, strength, and momentum |
| Paired task history | 200 seeds 2 histories 2 momentum protocols; 800 runs | A–B–C versus B–A–C after ten C epochs |
| Trajectory and events | Two cohorts of 200 seeds 2 histories; 800 runs | Earlier reduced states and one or four AB/BA events after five C epochs |
Appendix figures & tables29 assets
Supplementary material from the paper’s appendix.
Appendix
| Distance | Layer | Assignment | Gauge | MNIST | Fashion | KMNIST |
|---|---|---|---|---|---|---|
| Frob. | Both | Opt. | Canon. | 0.116 / 0.2935 | -0.181 / 0.8401 | 0.276 / 0.0490 |
| Frob. | Opt. | Canon. | 0.087 / 0.3641 | -0.410 / 0.9575 | 0.263 / 0.0682 | |
| Frob. | Opt. | Canon. | 0.054 / 0.3572 | 0.480 / 0.0007 | -0.002 / 0.5079 | |
| Frob. | Balanced | Opt. | Canon. | 0.030 / 0.4137 | 0.468 / 0.0009 | -0.001 / 0.5049 |
| Cosine | Both | Opt. | Canon. | 0.052 / 0.3581 | 0.057 / 0.3396 | 0.130 / 0.2085 |
| Cosine | Opt. | Canon. | 0.154 / 0.1421 | -0.073 / 0.6807 | -0.074 / 0.6873 |
| Dataset | Initial CV (%) | Changed (%) | Cycle E0 (%) | Cycle E20 (%) | Rank stability |
|---|---|---|---|---|---|
| MNIST | 0.700 | 92.05 | 14.70 | 26.93 | 1.0000 |
| Fashion | 0.751 | 93.51 | 15.77 | 28.98 | 1.0000 |
| KMNIST | 0.584 | 91.01 | 13.39 | 26.54 | 1.0000 |
| Analysis | E | Gauge | Layer | Norm | MNIST | Fashion | KMNIST |
|---|---|---|---|---|---|---|---|
| Distance | 0 | Raw | Sum | 0.115 / 0.2431 | 0.671 / 0.0004 | 0.371 / 0.0158 | |
| Distance | 0 | Raw | Difference | 0.143 / 0.1912 | -0.383 / 0.9701 | -0.110 / 0.7520 | |
| Distance | 20 | Raw | Sum | 0.179 / 0.2617 | -0.158 / 0.7208 | -0.164 / 0.7344 | |
| Distance | 20 | Raw | Difference | -0.021 / 0.5233 | 0.210 / 0.1464 | -0.157 / 0.7909 | |
| Partial RSA | 20 | Raw | Both | -0.045 / 0.6128 | 0.387 / 0.0033 | -0.192 / 0.9009 | |
| Distance | 0 | Raw | Sum | 0.132 / 0.1182 | 0.189 / 0.1961 | 0.333 / 0.0389 |
| Gauge | Layer | MNIST | Fashion | KMNIST |
|---|---|---|---|---|
| Raw | 0.261 / 0.2294 | -0.188 / 0.7118 | -0.442 / 0.9074 | |
| Raw | 0.418 / 0.1180 | 0.115 / 0.3750 | -0.527 / 0.9428 | |
| Raw | Both | 0.321 / 0.1811 | 0.152 / 0.3470 | 0.333 / 0.1753 |
| Canon. | 0.382 / 0.1407 | -0.006 / 0.5121 | -0.418 / 0.8930 | |
| Canon. | 0.382 / 0.1407 | -0.006 / 0.5121 | -0.418 / 0.8930 | |
| Canon. | Both | 0.382 / 0.1407 | -0.006 / 0.5121 | -0.418 / 0.8930 |
| Dataset | Retention | Chance | 95% interval | ||
|---|---|---|---|---|---|
| MNIST | 1 | 0.300 | 0.111 | [-0.366, 0.966] | 0.1248 |
| MNIST | 3 | 0.367 | 0.333 | [0.056, 0.678] | 0.4163 |
| Fashion | 1 | 0.200 | 0.111 | [-0.192, 0.592] | 0.3021 |
| Fashion | 3 | 0.233 | 0.333 | [-0.132, 0.598] | 0.9105 |
| KMNIST | 1 | 0.000 | 0.111 | [-0.320, 0.320] | 1.0000 |
| KMNIST | 3 | 0.433 | 0.333 | [0.059, 0.808] | 0.1595 |
| Dataset | Mean rank | Top-1 (%) | MRR | Holm |
|---|---|---|---|---|
| MNIST | 1.0 | 100 | 1.0 | 0.0003 |
| Fashion | 1.0 | 100 | 1.0 | 0.0003 |
| KMNIST | 1.0 | 100 | 1.0 | 0.0003 |
| Cost | Layer | Assignment | Gauge | MNIST | Fashion | KMNIST |
|---|---|---|---|---|---|---|
| Frob. | Native | Raw | 1.00 / 100 | 1.00 / 100 | 1.00 / 100 | |
| Frob. | Native | Raw | 1.00 / 100 | 1.00 / 100 | 1.00 / 100 | |
| Frob. | Both | Native | Raw | 1.00 / 100 | 1.00 / 100 | 1.00 / 100 |
| Unit Frob. | Native | Raw | 1.00 / 100 | 1.00 / 100 | 1.00 / 100 | |
| Unit Frob. | Native | Raw | 1.00 / 100 | 1.00 / 100 | 1.00 / 100 | |
| Unit Frob. | Both | Native | Raw | 1.00 / 100 | 1.00 / 100 | 1.00 / 100 |
| Dataset | Layer | Relative displacement | Exactly unchanged (%) | Change (%) |
|---|---|---|---|---|
| Fashion | 4.288 | 0.000 | 0.000 | |
| Fashion | 3.896 | 0.000 | 0.000 | |
| KMNIST | 5.053 | 0.000 | 0.001 | |
| KMNIST | 4.926 | 0.000 | 0.000 | |
| MNIST | 3.867 | 0.000 | 0.002 | |
| MNIST | 4.495 | 0.000 | 0.000 |
| Dataset | Features | Scale | Rank | 95% interval | Top-1 (%) | |
|---|---|---|---|---|---|---|
| MNIST | hidden | raw | 5.500 | [4.570, 6.430] | 0.0 | 0.5417 |
| MNIST | hidden | standardized | 5.300 | [4.418, 6.182] | 10.0 | 0.3568 |
| MNIST | logits | raw | 5.600 | [5.159, 6.041] | 10.0 | 0.6456 |
| Fashion | hidden | raw | 4.800 | [3.519, 6.081] | 10.0 | 0.0754 |
| Fashion | hidden | standardized | 5.200 | [4.840, 5.560] | 10.0 | 0.0950 |
| Fashion | logits | raw | 5.400 | [4.608, 6.192] | 20.0 | 0.4208 |
| Dataset | Space | Shared factor | Contrast | 95% interval | |
|---|---|---|---|---|---|
| MNIST | Weights | Init. | 0.1669 | [0.1560, 0.1777] | 0.0001 |
| MNIST | Weights | Schedule | 0.0019 | [0.0008, 0.0029] | 0.0031 |
| MNIST | CKA | Init. | 0.0446 | [0.0393, 0.0498] | 0.0001 |
| MNIST | CKA | Schedule | 0.0027 | [0.0013, 0.0042] | 0.0001 |
| Fashion | Weights | Init. | 0.0688 | [0.0584, 0.0793] | 0.0001 |
| Fashion | Weights | Schedule | 0.0074 | [0.0045, 0.0102] | 0.0001 |
| Dataset | Identify | Space | Rank range | Top-1 range (%) |
|---|---|---|---|---|
| Fashion | Init. | Weights | [1.025, 1.325] | [85.8, 98.3] |
| Fashion | Init. | CKA | [2.650, 2.925] | [41.7, 50.8] |
| Fashion | Schedule | Weights | [2.439, 2.589] | [30.8, 35.6] |
| Fashion | Schedule | CKA | [2.400, 2.506] | [32.5, 36.4] |
| KMNIST | Init. | Weights | [1.000, 1.000] | [100.0, 100.0] |
| KMNIST | Init. | CKA | [1.175, 1.358] | [85.0, 90.0] |
| Dataset | Measure | Mean | Init. fraction | Schedule fraction | Interaction/rest |
|---|---|---|---|---|---|
| MNIST | accuracy | 0.9727 | 0.181 | 0.166 | 0.654 |
| MNIST | loss | 0.1150 | 0.195 | 0.134 | 0.671 |
| Fashion | accuracy | 0.8638 | 0.111 | 0.438 | 0.451 |
| Fashion | loss | 0.4041 | 0.084 | 0.302 | 0.614 |
| KMNIST | accuracy | 0.8605 | 0.199 | 0.266 | 0.535 |
| KMNIST | loss | 0.7921 | 0.249 | 0.250 | 0.502 |
| E | Layer | (%) | : retained [CI] | : reset [CI] | Accuracy R/S | Region R/S |
|---|---|---|---|---|---|---|
| 2 | 1 | 60.29 [58.79, 61.58] | 59.89 [57.78, 61.92] | 97.36 / 97.30 | 90 / 90 | |
| 2 | 5 | 12.33 [11.95, 12.70] | 12.31 [11.92, 12.67] | 97.21 / 97.28 | 90 / 90 | |
| 2 | 10 | 6.29 [6.09, 6.48] | 6.32 [6.13, 6.52] | 97.33 / 97.32 | 80 / 80 | |
| 2 | 20 | 3.36 [3.27, 3.45] | 3.39 [3.30, 3.48] | 97.32 / 97.29 | 80 / 80 | |
| 2 | 1 | 59.89 [57.72, 61.94] | 58.87 [56.66, 60.78] | 97.27 / 97.26 | 90 / 80 | |
| 2 | 5 | 12.36 [11.96, 12.80] | 12.16 [11.87, 12.41] | 97.35 / 97.23 | 90 / 90 |
| E | Layer | (%) | : retained [CI] | : reset [CI] | Accuracy R/S | Region R/S |
|---|---|---|---|---|---|---|
| 2 | 1 | 68.55 [61.14, 75.60] | 72.31 [63.85, 80.60] | 86.75 / 86.70 | 60 / 70 | |
| 2 | 5 | 13.94 [12.33, 15.49] | 14.49 [12.61, 16.28] | 86.74 / 86.65 | 60 / 80 | |
| 2 | 10 | 7.27 [6.51, 8.05] | 7.38 [6.47, 8.27] | 86.81 / 86.75 | 50 / 80 | |
| 2 | 20 | 3.92 [3.48, 4.31] | 3.99 [3.61, 4.34] | 86.84 / 86.75 | 60 / 60 | |
| 2 | 1 | 67.42 [60.19, 74.06] | 70.39 [62.86, 78.57] | 86.84 / 86.67 | 70 / 80 | |
| 2 | 5 | 13.90 [12.26, 15.43] | 14.02 [12.59, 15.38] | 86.98 / 86.66 | 70 / 70 |
| E | Layer | (%) | : retained [CI] | : reset [CI] | Accuracy R/S | Region R/S |
|---|---|---|---|---|---|---|
| 2 | 1 | 78.85 [76.77, 80.90] | 79.15 [77.30, 80.83] | 85.97 / 85.98 | 90 / 90 | |
| 2 | 5 | 16.18 [15.79, 16.64] | 16.24 [15.92, 16.59] | 85.99 / 86.06 | 60 / 90 | |
| 2 | 10 | 8.22 [8.09, 8.35] | 8.37 [8.13, 8.62] | 86.19 / 86.13 | 100 / 90 | |
| 2 | 20 | 4.32 [4.25, 4.40] | 4.40 [4.31, 4.50] | 86.33 / 86.04 | 90 / 90 | |
| 2 | 1 | 77.99 [76.20, 79.80] | 78.58 [76.56, 80.83] | 86.25 / 86.26 | 100 / 100 | |
| 2 | 5 | 15.94 [15.58, 16.37] | 16.28 [15.83, 16.75] | 86.15 / 86.04 | 90 / 100 |
| Dataset | E | Returns | Accuracy (pp) | Loss | CKA range | Max. barrier |
|---|---|---|---|---|---|---|
| Fashion | 2 | 0 | [-0.25, 0.19] | [-0.0052, 0.0066] | [0.962, 0.970] | 0.00284 |
| Fashion | 10 | 8 | [-0.17, 0.11] | [-0.0060, 0.0068] | [0.974, 0.980] | 0.00361 |
| Fashion | 20 | 10 | [-0.25, 0.09] | [-0.0093, 0.0089] | [0.977, 0.984] | 0.00000 |
| KMNIST | 2 | 0 | [-0.19, 0.33] | [-0.0265, 0.0126] | [0.949, 0.960] | 0.00000 |
| KMNIST | 10 | 33 | [-0.14, 0.45] | [-0.0314, 0.0047] | [0.965, 0.969] | 0.00000 |
| KMNIST | 20 | 19 | [-0.38, 0.04] | [-0.0130, 0.0316] | [0.967, 0.970] | 0.00000 |
| Momentum | (%) | [95% CI] | Accuracy (%) | Loss | CKA | Local (%) |
|---|---|---|---|---|---|---|
| preserve | 1 | 16.134 [14.964, 17.294] | 97.599 | 0.12946 | 0.985 | 0 |
| reset | 1 | 15.646 [14.196, 16.916] | 97.610 | 0.12839 | 0.986 | 6 |
| preserve | 5 | 3.344 [3.121, 3.550] | 97.578 | 0.13079 | 0.985 | 0 |
| reset | 5 | 3.305 [3.056, 3.527] | 97.625 | 0.12880 | 0.986 | 0 |
| preserve | 20 | 1.165 [1.128, 1.202] | 97.612 | 0.13113 | 0.984 | 0 |
| reset | 20 | 1.161 [1.120, 1.196] | 97.612 | 0.12998 | 0.984 | 0 |
| Momentum | (%) | q95 | q90 | Half A | Half B |
|---|---|---|---|---|---|
| preserve | 1 | 72 [48, 92] | 72 [48, 92] | 68 [42, 90] | 52 [30, 74] |
| reset | 1 | 74 [52, 92] | 74 [52, 92] | 64 [40, 84] | 50 [28, 74] |
| preserve | 5 | 62 [38, 84] | 58 [34, 80] | 56 [34, 78] | 50 [28, 72] |
| reset | 5 | 76 [52, 96] | 72 [46, 94] | 72 [46, 94] | 60 [42, 76] |
| preserve | 20 | 34 [12, 58] | 32 [10, 56] | 34 [10, 60] | 42 [18, 66] |
| reset | 20 | 42 [22, 64] | 36 [16, 60] | 44 [18, 72] | 52 [26, 78] |
| (%) | Measure | Difference | 95% interval |
|---|---|---|---|
| 1 | ratio | 0.48823 | [-0.25730, 1.26009] |
| 1 | distance | 0.08921 | [-0.04705, 0.22988] |
| 1 | accuracy | -0.00011 | [-0.00043, 0.00018] |
| 1 | loss | 0.00107 | [-0.00038, 0.00231] |
| 1 | region member | -0.02000 | [-0.12000, 0.06000] |
| 1 | local member | -0.06000 | [-0.18000, 0.00000] |
| (%) | Runs | Distance [95% CI] | Accuracy [95% CI], % | Loss [95% CI] |
|---|---|---|---|---|
| 0 | 10 | 3.318 [3.132, 3.506] | 97.56 [97.44, 97.70] | 0.1336 [0.1269, 0.1397] |
| 20 | 50 | 4.457 [4.328, 4.591] | 97.58 [97.51, 97.66] | 0.1320 [0.1278, 0.1360] |
| Stage | Assignment | Cells | Below 1 | Above 1 | Includes 1 |
|---|---|---|---|---|---|
| Expanded | Git Re-Basin | 216 | 0 | 216 | 0 |
| Expanded | PIF/Hungarian | 216 | 0 | 216 | 0 |
| Expanded | Alpha matching | 216 | 0 | 215 | 1 |
| Pilot | Git Re-Basin | 6 | 0 | 6 | 0 |
| Pilot | PIF/Hungarian | 6 | 0 | 6 | 0 |
| Pilot | Alpha matching | 6 | 0 | 6 | 0 |
| Momentum | C epochs | Features | Accuracy | 95% interval | Holm | |
|---|---|---|---|---|---|---|
| reset | 1 | 100.00 | [100.00, 100.00] | 0.001 | 0.019 | |
| reset | 1 | 100.00 | [100.00, 100.00] | 0.001 | 0.019 | |
| reset | 1 | Both weights | 100.00 | [100.00, 100.00] | 0.001 | 0.019 |
| reset | 1 | Diagnostics | 85.00 | [77.50, 92.50] | 0.001 | 0.019 |
| reset | 1 | Logits | 96.25 | [91.25, 100.00] | 0.001 | 0.019 |
| reset | 1 | Residual weights | 83.75 | [75.00, 91.25] | 0.001 | 0.019 |
| Momentum | Measure | Difference | 95% interval |
|---|---|---|---|
| reset | accuracy | -0.8065 | [-1.1788, -0.4812] |
| reset | loss | 0.0246 | [0.0124, 0.0374] |
| continued | accuracy | -0.8170 | [-1.1280, -0.5179] |
| continued | loss | 0.0275 | [0.0171, 0.0380] |
| Cohort | Layer | PCA dim. | Probe | MSE | MSE 95% interval | |
|---|---|---|---|---|---|---|
| One-pair | Both | 8 | RIDGE | -0.199 | 0.06937 | [0.06189, 0.07681] |
| One-pair | Both | 16 | RIDGE | -0.079 | 0.06622 | [0.06170, 0.07029] |
| One-pair | Both | 32 | RIDGE | -0.040 | 0.06229 | [0.05921, 0.06521] |
| One-pair | Both | 8 | MLP | -1.869 | 0.16599 | [0.14798, 0.18650] |
| One-pair | 8 | RIDGE | -0.200 | 0.06940 | [0.06191, 0.07684] | |
| One-pair | 16 | RIDGE | -0.080 | 0.06627 | [0.06174, 0.07033] |
| Cohort | Coordinates | Explained variance (%) |
|---|---|---|
| One-pair | 8 | 27.57 |
| One-pair | 16 | 32.07 |
| One-pair | 32 | 38.14 |
| Four-pair | 8 | 28.33 |
| Four-pair | 16 | 33.09 |
| Four-pair | 32 | 39.85 |
| Cohort | Probe | 95% interval | MSE | |
|---|---|---|---|---|
| One-pair | Ridge | -0.0272 | [-0.0762, -0.0166] | 0.30868 |
| One-pair | Mean | -0.0155 | [-0.0595, -0.0097] | 0.30516 |
| Four-pair | Ridge | -0.0082 | [-0.0484, -0.0018] | 0.09759 |
| Four-pair | Mean | -0.0143 | [-0.0543, -0.0099] | 0.09819 |
| Cohort | Position | Features | Probe | Accuracy (%) | 95% interval | Holm |
|---|---|---|---|---|---|---|
| One-pair | 1 | Both | Ridge | 100.00 | [100.00, 100.00] | 0.006 |
| One-pair | 1 | Both | Shuffled | 55.00 | [38.75, 68.78] | – |
| One-pair | 1 | Both | MLP | 100.00 | [100.00, 100.00] | – |
| One-pair | 1 | W0 | Ridge | 100.00 | [100.00, 100.00] | 0.006 |
| One-pair | 1 | W1 | Ridge | 100.00 | [100.00, 100.00] | 0.006 |
| One-pair | 1 | performance | Ridge | 70.00 | [61.25, 78.75] | 0.006 |
| Cohort | Split | Position | Measure | Difference | 95% interval |
|---|---|---|---|---|---|
| One-pair | all | 1 | accuracy | -1.086 | [-1.309, -0.877] |
| One-pair | all | 1 | loss | 0.036 | [0.029, 0.043] |
| One-pair | test | 1 | accuracy | -0.949 | [-1.341, -0.564] |
| One-pair | test | 1 | loss | 0.034 | [0.020, 0.049] |
| Four-pair | all | 1 | accuracy | -0.459 | [-3.924, 3.010] |
| Four-pair | all | 1 | loss | 0.015 | [-0.064, 0.092] |
| Cohort | Position | Successful pairs | Rate (%) | Exact 95% interval |
|---|---|---|---|---|
| One-pair | 1 | 40/40 | 100.0 | [91.19, 100.00] |
| Four-pair | 1 | 35/40 | 87.5 | [73.20, 95.81] |
| Four-pair | 2 | 17/40 | 42.5 | [27.04, 59.11] |
| Four-pair | 3 | 16/40 | 40.0 | [24.86, 56.67] |
| Four-pair | 4 | 14/40 | 35.0 | [20.63, 51.68] |
| Four-pair | All four | 1/40 | 2.5 | [0.06, 13.16] |
Neural Network Training Dynamics
Neural network training dynamics studies how parameters and representations change during optimization and how those changes shape learning, stability, and generalization.
See all 553 papers