Changelog

Major update with expanded flexibility, improved diagnosis tools, and uncertainty quantification.
Additional distribution families: now supports poisson in addition to gaussian and binomial.
Per-term architecture configuration: hyperparameters (units, activation, learning rate, initializers, regularizers) can now be set per smooth term inside s().
Confidence Intervals (CI):
- uncertainty_method argument allows estimation of epistemic uncertainty.
- Intervals integrated into predict() and autoplot().
Cross-validation support: new validation_split parameter for monitoring validation losses during training.
Training diagnostics: new plot_history() function for visualizing training/validation loss curves per term and per backfitting iteration.
Improved summary(): displays per-term configuration, layer architectures, linear coefficients, and compact training history.
Diagnosis plots: new diagnose() function which provides a 2×2 diagnostic panel.
Autoplot enhancements: ggplot2-based diagnostic and effect plots with support for CI ribbons, per-term inspection, and factor vs continuous term visualization.
Testing: expanded test coverage for new families, CI estimation, plotting, and per-term configuration.
Internal refactoring:
- Clean separation of deviance and link functions per family.
- Consistent handling of sample weights.
- Improved numerical stability (clamping in log/exp/probabilities).

verbose parameter is now used along all the required functions.
Tensorflow and Keras are now loaded when library(neuralGAM) is invoked for the first time, and therefore the first run of the neuralGAM() function has all the required packages ready.

neuralGAM 2.0