Datasets:
|
Download reports/Online_Learning.md from PureOne/Iyeumaring-AION-LATCH-VI: direct link, hf CLI and curl.
- Browser
- Download file 9.84 kB
-
https://huggingface.co/datasets/PureOne/Iyeumaring-AION-LATCH-VI/resolve/main/reports/Online_Learning.md
- Command line
-
hf download hf://datasets/PureOne/Iyeumaring-AION-LATCH-VI/reports/Online_Learning.md
-
curl -L -o Online_Learning.md https://huggingface.co/datasets/PureOne/Iyeumaring-AION-LATCH-VI/resolve/main/reports/Online_Learning.md
9.84 kB
| # Certificate-monotone online learning and acquisition controls | |
| ## What is demonstrated | |
| A finite-label learner can update repeatedly while preserving the protected-domain, residual-Lipschitz and head-Lipschitz contracts from the scalar filter construction. With ordinary maximum-uncertainty acquisition and head cap C=2, increasing labels from 8 to 64 reduces the mean held-out pointwise error certificate from 0.17931 to 0.10685 and relative test MSE from 0.52504 to 0.24979. The system learns an unchanged unknown residual from additional observations. This is online active learning with analytic old-domain retention, not autonomous RSI or a sequence of newly invented learning algorithms. | |
| The experiment also implements a new explicit acquisition objective based on guaranteed post-query certificate reduction. Its mathematical bound is valid and nontrivial, but it does not beat the strong uncertainty baseline on held-out certificates and costs substantially more computation. Both findings are retained. | |
| ## Incremental decoder and monotonicity | |
| The supplied contracts are exact protected geometry K, distance D, residual Lipschitz bound L=1 and observation error at most sigma=0.01. The target is scalar. Maintain the finite-data lower and upper envelopes l,u from the previous learning study. A new observed label y at x_i updates them by | |
| l_new(x)=max(l(x), y-sigma-L||x-x_i||), | |
| u_new(x)=min(u(x), y+sigma+L||x-x_i||). | |
| This takes O(qd) work for q query points when distances are not cached, or O(q) with a cached distance column. It requires no recomputation over every earlier label. The intervals remain nested and contain the correctly specified unknown residual. Their midpoint remains L-Lipschitz and zero on K. Each deployed learner replaces the previous prediction with the fixed nonlinear one-quarter filter applied to the new midpoint; it does not accumulate residual adapters and their Lipschitz constants. | |
| At fixed D and C, let p_D(z) be the filtered prediction and b_D(z)=|z-p_D(z)|. For the implemented filter, b is even, nondecreasing in |z| and 1-Lipschitz. The exact pointwise conditional worst-case absolute error is | |
| R(l,u)=(u-l)/2+b_D((l+u)/2). | |
| If I_new is a subinterval of I, its radius decrease is at least its midpoint displacement. The 1-Lipschitz property of b therefore gives R_new<=R. This certificate is nonincreasing after every consistent observation, not merely on average. D, C and the noise/Lipschitz contracts must remain fixed; the proof does not automatically cover changing geometries, changing caps, contradictory labels or arbitrary target drift. | |
| ## A query with a guaranteed value | |
| At the query point itself, an observation has noise radius sigma. If u-l<=2sigma, some feasible observation can leave the entire prior interval unchanged. Otherwise the worst possible posterior certificate is exactly | |
| sigma + max[b_D(l+sigma), b_D(u-sigma)]. | |
| Subtracting this from the current R gives a nonnegative guaranteed local certificate reduction. The formula follows by intersecting the interval with the observed noise tube: worst posterior risk occurs at a full-width tube at one of the two interval edges. | |
| Querying x_i also supplies a valid tube at another pool point x_j with effective radius s_ij=sigma+L||x_i-x_j||. Apply the same worst-posterior formula at j, using s_ij, and sum its guaranteed reductions over j. This is a lower bound on the actual decrease of the mean pool-point certificate for every feasible observed label. Separate worst labels may differ across j, so the sum is conservative. Maximizing it is not claimed to solve globally optimal experimental design or to minimize held-out MSE. | |
| ## Frozen practical comparison | |
| The source and protocol were frozen before 16 new seeds, 80000 through 80015. Source SHA-256: | |
| d462cb6b2309e016bdf53d1e1b17bbd8bfdfba3baf52205e0620aa580767927a | |
| There are three dimension/plane configurations (1/1,2/2,4/4), four target families (signed distance, distance, smooth product and oscillatory), head caps 2 and 8, and seven acquisition rules. A pool has 128 possible observations. Every method starts from the same eight labels and reports checkpoints at 8,16,32 and 64 labels. The four rounds do not change the learning rule. There are 2,688 episodes and 10,752 checkpoint rows. | |
| Controls are random sampling, maximum geometric separation, maximum interval radius, maximum current filtered risk, maximum local guaranteed risk reduction, a strong coverage-radius rule, and the global guaranteed-risk rule. Coverage-radius maximizes sum_j[max(0,radius_j-sigma-L distance_ij)], a guaranteed radius-improvement bound. All rules use the same decoder, data, initial observations and label budget. Query selection never reads unobserved pool labels or test targets. The direct unfiltered midpoint is also evaluated as a stronger prediction/certificate control that has no common head-regularity guarantee. | |
| | Cap | Rule | Certificate at 8 | Certificate at 64 | Relative test MSE at 64 | Cold episode ms | | |
| |---:|---|---:|---:|---:|---:| | |
| | 2.0 | random | 0.17931 | 0.11523 | 0.29346 | 6.32 | | |
| | 2.0 | maximin | 0.17931 | 0.10971 | 0.27227 | 6.07 | | |
| | 2.0 | max_radius | 0.17931 | 0.10685 | 0.24979 | 6.15 | | |
| | 2.0 | max_current_risk | 0.17931 | 0.10782 | 0.25011 | 7.23 | | |
| | 2.0 | local_risk_drop | 0.17931 | 0.10725 | 0.24838 | 9.29 | | |
| | 2.0 | coverage_radius | 0.17931 | 0.10724 | 0.25034 | 7.88 | | |
| | 2.0 | global_risk_drop | 0.17931 | 0.10731 | 0.24869 | 31.05 | | |
| | 8.0 | random | 0.16304 | 0.09133 | 0.27280 | 6.24 | | |
| | 8.0 | maximin | 0.16304 | 0.08514 | 0.25087 | 6.15 | | |
| | 8.0 | max_radius | 0.16304 | 0.08167 | 0.22826 | 6.10 | | |
| | 8.0 | max_current_risk | 0.16304 | 0.08189 | 0.22827 | 7.32 | | |
| | 8.0 | local_risk_drop | 0.16304 | 0.08166 | 0.22795 | 9.23 | | |
| | 8.0 | coverage_radius | 0.16304 | 0.08208 | 0.22875 | 7.87 | | |
| | 8.0 | global_risk_drop | 0.16304 | 0.08202 | 0.22832 | 30.64 | | |
| For C=2 maximum-radius sampling, the 8-to-64 label certificate difference is -0.07246 with 95% paired seed-bootstrap interval [-0.07794,-0.06802]. The MSE difference is -0.27525, interval [-0.29912,-0.25318]. These improvements demonstrate continued data-driven learning under the fixed class contract. MSE itself has no monotonicity theorem. | |
| The global guaranteed-risk rule's final C=2 certificate is worse than maximum radius by 0.000468, interval [0.000348,0.000585]. Its test MSE is better by only 0.00110, interval [-0.00182,-0.00045], while its cold cost is about five times larger. The standard maximum-radius rule is the better practical choice on this suite. These descriptive intervals use 5,000 paired resamples of 16 seed means, averaging the 12 fixed geometry/family cases; no multiplicity adjustment is applied. | |
| ## Bounds, costs and limits | |
| No pointwise certificate increases occur at any observed update, and no held-out error certificate is violated. Every one of the 10,752 post-initial queries per cap under the global rule has a strictly positive predicted pool-risk drop. At C=2 the mean cumulative guaranteed drop after the initial labels is 0.07581 versus actual pool drop 0.10879; at C=8 the values are 0.10164 and 0.12288. Maximum guarantee shortfall is 4.26e-17, a floating-point effect. These are pool-certificate bounds, not bounds on held-out MSE or the realized value of an expensive external experiment. | |
| Old-plane update leakage is zero in the recorded checkpoints. Maximum measured head-slope/cap ratio is 1+2.57e-12 and maximum residual slope 1+8.22e-11. Finite-difference rounding explains those tiny overshoots; the analytic class theorem, rather than a finite grid, supplies the global statement. The direct midpoint's measured head-slope/cap ratio reaches 81.33, showing why its better conditional-risk value cannot be substituted without checking the head contract. | |
| All policies use a shared full geometric pool/query distance cache. Cold times charge its construction plus the measured policy episode, including updates and checkpoint predictions. The simulation actually reuses the cache across comparator policies; adding its measured construction cost gives a counterfactual cold cost per policy. There are 704 query/probe points. The global/coverage scores are O(pool^2) per label; maximum-radius scores are O(pool). Lazy distance construction could change absolute costs and was not implemented. Real label-acquisition latency, data acquisition and synthetic target generation are excluded. No runtime leadership is claimed for the new objective. | |
| The learner uses correctly specified known L, noise and protected geometry on one fixed residual task. It does not estimate those contracts, demonstrate arbitrary distribution drift, or learn a different training algorithm. High-dimensional coverage and finite-pool generalization remain separate limitations. The nontrivial result is that a real label stream drives repeatedly improving uncertainty certificates while every deployed predictor stays in the protected head class. | |
| `online_certificates.py` contains the complete source. `protocol_confirmatory.json` records the freeze. `confirmatory/episodes.json` stores every selected query and the per-query guaranteed/actual progress trace; `rows.csv`, `summary.json` and `paired_analysis.json` preserve all comparisons. Three test groups check the local minimax formula against a dense label grid, monotone updates and the global pool bound. Independent external replication has not occurred. | |
| Reproduce with Python 3.12 and NumPy 2.3.5: | |
| OPENBLAS_NUM_THREADS=1 OMP_NUM_THREADS=1 python \ | |
| research/continuation_v6/learning/online_certificates.py --test | |
| OPENBLAS_NUM_THREADS=1 OMP_NUM_THREADS=1 python \ | |
| research/continuation_v6/learning/online_certificates.py --seeds 16 \ | |
| --out research/continuation_v6/learning/reproduced | |
| Matplotlib is needed only for the chart. | |