Released Jan. 31st, 2026.

• Allow batching when training RNN with ragged lengths (#4042).

• Fixed generated artifiacts in resized float images by using clamping (#4030).

• Optimize convolution (#3988).

• Added GELUExact ANN activation layer (#3994).

• Adapt GRU ANN layer to the new interface (#3955).

• Fix warning on CRAN for bundled STB (#3950).

• Fix potential MSVC constructor shadowing (#3958).

• Use a looser check for auto-detection of categorical file types (#3961).

• Accelerate CELU layer (#3975).

• Fix dependency detection bugs in mlpack.cmake (#3981).

• Add a SumReduce layer (#3991).

• Update header used by R packages compiling directly against C++ API (#3990).

• Add FFN::Add() and RNN::Add() with copy and move semantics; use these instead of passing layers to Add() as pointers (#3974).

• Mark long-running tests with the [long] tag (#3983).

• Added DAGNetwork class to represent complex neural network structures (#3944).

• Fix mask handling in MultiHeadAttention layer (#3998)

• Added data::GroupChannels() and data::InterleaveChannels() for preprocessing images before using Convolution layers. (#4006)

• Fix infinite recursion in Octree when the number of identical points exceeds maxLeafSize (#4020).

• Add Embedding layer (#3999).

• Add YOLOv3Tiny for object detection (#4023).

• Fix thread-specific random seed initialization (#4027).

• R bindings now return class attributes as vector with the given model, mlpack model bindings and list as fallback (#4045).

• Added BoundingBoxImage() for drawing bounding boxes onto images when doing tasks such as object detection (#4039).

Released May 22, 2025.

• Fix compilation of Save() when HDF5 is enabled (#3942).

• Update bundled STB to fix warnings in R bindings (#3940).

• Fix cross-validation support for algorithms with many parameters (including RandomForest) (#3941).

Released May 14, 2025.

• Shuffle sequence lengths for RNNs (#3926).
• Add ability to compile OpenBLAS for Windows (#3922).
• Drop pytest-runner and "setup.py test" support (#3921).
• Fix compilation errors with clang++ version 20 (#3928).

Released Apr. 3, 2025.

• Fix command-line duplicate output bug when loading matrices for some bindings (#3838).
• Use CMAKE_BUILD_TYPE to specify build type instead of DEBUG and PROFILE options (#3865).
• Add MLPACK_NO_STD_MUTEX to allow disabling std::mutex (#3868).
• Bundle STB with mlpack and add ResizeImages() functionality (#3823).
• Add mlpack.cmake to facilitate finding mlpack and its dependencies (#3872).
• Fix conversion of empty Armadillo objects to numpy in Python bindings (#3896).
• Added bootstrap strategies for RandomForest: IdentityBootstrap, DefaultBootstrap, and SequentialBootstrap (#3829).
• Add ResizeCropImages() for resize-and-crop image preprocessing functionality (#3903).
• Fix LSTM input size calculation for multidimensional inputs (#3913).

Released Dec. 4, 2024.

• Fix compilation with clang 19 (#3799).
• Deprecate version of data::Split() that returns a std::tuple for consistency; use other overloads instead (#3803).
• Fix LSTM layer copy/move constructors (#3809).
• Fix compilation if only including mlpack/methods/kde/kde_model.hpp (#3800).
• Fix serialization and MinDistance() bugs with HollowBallBound (#3808).

Released September 18, 2024.

• Distribute STB headers as part of R package (#3724, #3726).
• Added OpenMP parallelization to Hamerly, Naive, and Elkan k-means (#3761, #3762, #3764).
• Added OpenMP support for fast approximation (#3685).
• Implemented the Find and Fill algorithm into the Dropout Layer and added OpenMP support (#3684).
• Update Python bindings to support NumPy 2.x (#3752).
• Bump minimum Armadillo version to 10.8 (#3760).
• Adapt NearestInterpolation ANN layer to new Layer Interface (#3768).
• Add support for arbitrary matrix types to Radical and deprecate Radical::DoRadical() in favor of Radical::Apply() (#3787).

Released May 28, 2024.

• Add print_training_accuracy option to LogisticRegression bindings (#3552).
• Fix preprocess_split() call in documentation for LinearRegression and AdaBoost Python classes (#3563).
• Added Repeat ANN layer type (#3565).
• Remove round() implementation for old MSVC compilers (#3570).
• (R) Added inline plugin to the R bindings to allow for other R packages to link to headers (#3626, h/t @cgiachalis).
• (R) Removed extra gcc-specific options from Makevars.win (#3627, h/t @kalibera).
• (R) Changed roxygen package-level documentation from using @docType package to "_PACKAGE". (#3636)
• Fix floating-point accuracy issue for decision trees that sometimes caused crashes (#3595).
• Use templates metaprog to distinguish between a matrix and a cube type (#3602), (#3585).
• Use MatType instead of arma::Mat<eT>, (#3567), (#3607), (#3608), (#3609), (#3568).
• Generalize matrix operations for armadillo and bandicoot, (#3619), (#3617), (#3610), (#3643), (#3600), (#3605), (#3629).
• Change arma::conv_to to ConvTo using a local shim for bandicoot support (#3614).
• Fix a bug for the stddev and mean in RandNormal() #(3651).
• Allow PCA to take different matrix types (#3677).
• Fix usage of precompiled headers; remove cotire (#3635).
• Fix non-working verbose option for R bindings (#3691), and add global mlpack.verbose option (#3706).
• Fix divide-by-zero edge case for LARS (#3701).
• Templatize SparseCoding and LocalCoordinateCoding to allow different matrix types (#3709, #3711).
• Fix handling of unused atoms in LocalCoordinateCoding (#3711).
• Move minimum required C++ version from C++14 to C++17 (#3704).

Released Nov. 27, 2023.

• Fix include ordering issue for LinearRegression (#3541).
• Fix L1 regularization in case where weight is zero (#3545).
• Use HTTPS for all auto-downloaded dependencies (#3550).
• More robust detection of C++17 mode in the MSVC "compiler" (#3555, #3557).
• Fix setting number of classes correctly in SoftmaxRegression::Train() (#3553).
• Adapt MultiheadAttention and LayerNorm ANN layers to new Layer interface (#3547).
• Fix inconsistent use of the "input" parameter to the Backward method in ANNs (#3551).

Released Sep. 7, 2023. (Sorry for the late Github release. Forgot to hit the "publish" button.)

• Reinforcement Learning: Gaussian noise (#3515).
• Reinforcement Learning: Twin Delayed Deep Deterministic Policy Gradient (#3512).
• Reinforcement Learning: Ornstein-Uhlenbeck noise (#3499).
• Reinforcement Learning: Deep Deterministic Policy Gradient (#3494).
• Add ClassProbabilities() member to DecisionTree so that the internal details of trees can be more easily inspected (#3511).
• Bipolar sigmoid activation function added and invertible functions fixed (#3506).
• Add auto-configured mlpack/config.hpp to contain configuration details of mlpack that are required at compile time. STB detection is now done in this file with the MLPACK_HAS_STB macro (#3529).

Released June 16, 2023.

• Adapt C_ReLU, ReLU6, FlexibleReLU layer for new neural network API (#3445).
• Fix PReLU, add integration test to it (#3473).
• Fix bug in LogSoftMax derivative (#3469).
• Add serialize method to GaussianInitialization, LecunNormalInitialization, KathirvalavakumarSubavathiInitialization, NguyenWidrowInitialization, and OrthogonalInitialization (#3483).
• Allow categorical features to preprocess_one_hot_encode (#3487).
• Install mlpack and cereal headers as part of R package (#3488).
• Add intercept and normalization support to LARS (#3493).
• Allow adding two features simultaneously to LARS models (#3493).
• Adapt FTSwish activation function (#3485).
• Adapt Hyper-Sinh activation function (#3491).