NEWS

---

neuralnetwork 0.1.4 (2026-10-10)

This update includes the unreleased 0.1.2 and 0.1.3 corrections. Users upgrading from 0.1.1 should recompute classification metrics and refit models affected by the weighted training, formula preprocessing or checkpoint-selection fixes.

Training and numerical stability

• Fixed weighted mini-batch gradients: weights no longer cancel in one-row batches. The normalization now matches the full weighted objective.
• Excluded zero-weight training rows from optimization, scalers and fitted formula bases. Fitted metadata records how many training rows were excluded.
• Finite gradient clipping now selects Adam automatically; explicit L-BFGS rejects the unsupported control instead of silently ignoring it.
• Shared L-BFGS loss/gradient forward passes and allocated optimizer state only when needed. These changes do not add runtime dependencies.
• Avoided cancellation in near-hard binary soft-target loss, and intermediate overflow/underflow in extreme multiclass loss and L2 gradients.
• Preserved RPROP direction-change detection for extremely small gradients by comparing signs rather than multiplying gradient magnitudes.
• Clarified errors for complex input and hidden-layer integers outside the supported range; rejected non-finite scaling before native kernels.
• Added deep-network gradient and weighted batch-average oracles, one-row weight checks, zero-weight invariance, cache instrumentation and allocation tests. Prior numerical and public-interface tests remain in the suite.
• Wrapped reference examples and help-page text for wider code fonts as well as the macOS manual font. No API defaults changed.

Evaluation, formulas and model selection

• Corrected multi-output R-squared centering and aligned training RMSE with evaluation RMSE. Constant truth returns an undefined R-squared explicitly.
• Preserved fitted formula bases and contrasts, applied response transformations during evaluation, and learned holdout bases from training rows only.
• Rejected mismatched response rows and missing named predictor columns; aligned named outcome matrices and accepted classification indicator/soft-target truth.
• Added sample-weighted evaluation and importance. Cross-validation now splits weights, stratifies encoded classes, retains undefined folds and reports score counts. All helpers use the fitted unpenalized data loss for metric = "loss".
• Tuning now compares candidates on one common holdout and ranks only finite scores. Candidate identifiers preserve the link to stored models.
• Used stable logits for classification objectives so saturated losses agree with gradients. Fixed the mini-batch L2 denominator, skipped zero-weight batches, rejected non-finite optimizer states and separated checkpoint saving from the patience improvement threshold.
• Removed fabricated intervals from singular Hessians. Parameter intervals are restricted to identifiable unweighted, unregularized no-hidden-layer linear or binary logistic L-BFGS fits, using total information and proper dispersion.
• Corrected sensitivities to use pre-scaling input units, fixed intercept-only parameter serialization and single-row permutations, and validated saved model structure on load. Unsupported L-BFGS epoch controls now fail explicitly.
• Added independent gradient, optimizer, regression, formula, selection and inference tests; tightened the release checker to reject any unapproved NOTE.
• Corrected plotting labels, included both loss series in the vertical range, drew single stored diagnostics as points, and stopped sensitivity plots from magnifying numerical noise in nearly constant derivatives.

Missed-class metrics

• Fixed inflated macro F1 and macro precision when a fitted class has no predicted observations. F1 now uses confusion-matrix counts directly, so wholly missed classes also contribute zero. Thanks to Graham Upton for the reproducible report.
• Macro precision, recall, and F1 now average over all fitted classes, using zero for a class-wise zero denominator. Binary metrics use the same rule; balanced accuracy still averages recall over classes present in the truth.
• Added regression tests for the reported matrix, 591 exhaustive binary and multiclass tables, and both prediction backends and model-selection paths.
• Made binary model-selection scores agree with prediction at a probability of exactly 0.5, and reject empty classification evaluations with a clear error.
• Documented the class averaging and zero-denominator rules with a worked confusion-matrix example.
• Clarified that printed fit metrics describe the selected checkpoint, not necessarily the last epoch.

neuralnetwork 0.1.1 (2026-09-07)

Maintenance release.

• Added optimizer convergence metadata to fitted models. L-BFGS fits now expose the convergence code and message returned by stats::optim(), and summary() prints that diagnostic.
• Added an explicit stratified cross-validation guard: when stratify = TRUE, each class must have at least k rows. This fails early with a clear message instead of producing a later fold-specific class mismatch.
• Added error_action = "continue" to nn_tune() so exploratory grids can record failed candidates and continue ranking the usable fits.
• Added explicit outcome-length checks for evaluation, permutation importance, and finite-difference Hessian helpers.
• Improved prediction errors for missing predictors and factor levels not seen during fitting.
• Added validation for classification target matrices and for missing or empty inputs to nn_class_ind().
• Added explicit validation for finite-difference step sizes, Hessian parameter limits, and confidence levels.
• Added stricter validation for optimizer hyperparameters, random seeds, integer loop counts, and callback return values.
• Clarified L-BFGS printed output so it reports function evaluations rather than epochs.
• Hardened the local CRAN check script so it parses R CMD check status lines and treats ERROR/WARNING results as failures.
• Added tests and documentation for the new diagnostics and cross-validation validation rule.

neuralnetwork 0.1.0 (2026-06-20)

First CRAN-oriented release.

• Added nn_fit() for compact multilayer perceptrons with formula, data frame, matrix, and vector inputs.
• Added regression, binary classification, and multiclass classification.
• Added Adam, SGD, momentum, Nesterov, RPROP, GRPROP, and L-BFGS optimizers.
• Added automatic hidden-layer sizing, optimizer selection, and activation selection.
• Added optional portable Rcpp forward-pass kernels.
• Added dropout, L2 regularization, gradient clipping, learning-rate decay, validation splits, early stopping, and callback hooks.
• Added sample weights and balanced class weights.
• Added robust Huber loss for regression.
• Added task-aware evaluation metrics, tuning, repeated k-fold cross-validation, and permutation importance.
• Added save/load helpers and S3 methods for prediction, printing, plotting, summaries, and coefficients.
• Added compatibility helpers for common nnet and neuralnet tasks.
• Added a worked-example vignette.