R2026a

New Features, Bug Fixes, Compatibility Considerations
expand all

Deployment

 Generate C/C++ code and deploy pretrained XGBoost models (requires MATLAB Coder)

You can generate C/C++ code and perform inference using pretrained XGBoost models. Save a CompactClassificationXGBoost or CompactRegressionXGBoost model object by using saveLearnerForCoder, and then load the model in the entry-point function for code generation by using loadLearnerForCoder.

Anomaly Detection Blocks: Find anomalies in data and generate code in Simulink

You can integrate the isanomaly function of isolation forest objects and isanomaly function of the one-class SVM objects into Simulink® using the new anomaly detector blocks. Import a trained anomaly detection model and configure data types using the Block Parameters dialog box.

The blocks support C/C++ code generation and fixed-point conversion. You can generate C and C++ code using Simulink Coder™, and design and simulate fixed-point systems using Fixed-Point Designer™.

Override model properties in the IncrementalClassificationNaiveBayes Predict Block

You can override model properties such as Cost, Prior, and ScoreTransform in the IncrementalClassificationNaiveBayes Predict block using the Block Parameters dialog box. In previous releases, you set these properties when creating the initial incrementalClassificationNaiveBayes model object.

ClassificationECOC Predict block: Use machine learning models trained with binary kernel learners

The ClassificationECOC Predict block now supports ClassificationECOC and CompactClassificationECOC machine learning models trained with binary kernel learners.

Generate C/C++ code for performing incremental learning using kernel classification and regression functions (requires MATLAB Coder)

You can generate C/C++ code that trains a kernel classification or regression model in real time, as batches of observations become available. You can save an incrementalClassificationKernel or incrementalRegressionKernel model by using saveLearnerForCoder, and then load the model in the entry-point function for code generation by using loadLearnerForCoder (see Introduction to Code Generation for Statistics and Machine Learning Functions). These incremental learning functions support code generation:

  • The fit function fits the model to incoming data by updating the coefficients.

  • The predict function predicts responses given incoming data.

  • The loss function returns the classification or regression loss of the model for incoming data.

  • The updateMetrics function updates performance metrics given incoming data.

  • The updateMetricsAndFit function updates performance metrics and fits the model to incoming data.

Generate C/C++ code for performing incremental learning using a multiclass classification model with kernel binary learners (requires MATLAB Coder)

You can generate C/C++ code for incrementalClassificationECOC model objects that use kernel binary learners. For more information, see Extended Capabilities and Introduction to Code Generation for Statistics and Machine Learning Functions.

saveLearnerForCoder Function: Specify kernel computation method for Gaussian SVM and ECOC models

If your trained model is a Gaussian SVM model, or an ECOC model with at least one Gaussian SVM binary learner, you can specify OptimizeKernelFor=ComputationMethod in saveLearnerForCoder to select the kernel computation method to use in the generated C/C++ code. When you specify OptimizeKernelFor="speed" (the default), the SVM model uses an expansion formulation for fast calculations of the Gram matrix. In some cases, such as when your trained model has a small kernel scale, this formulation might lead to numerical precision loss. To use an exact formulation for the calculations, specify OptimizeKernelFor="accuracy". This setting typically results in slower calculations.

pdist and pdist2 Functions: Generate optimized CUDA code using Mahalanobis distance metric (requires GPU Coder)

The pdist and pdist2 functions now support optimized CUDA® code generation for the Mahalanobis distance metric when you specify the Distance input argument as "mahalanobis".

 Multithreading support for C/C++ code generation of ensembles of trees and ensembles of discriminant analysis models (requires MATLAB Coder)

When you generate single- or double-precision C/C++ code for ensembles consisting of all tree learners or all discriminant analysis learners, the generated code of the predict object function now uses parfor to create loops that run in parallel on supported shared-memory multicore platforms. For more information, see Code Generation (for classification) and Code Generation (for regression).

Apps

 Machine Learning Apps: Import a trained model from the MATLAB workspace

Import a supported trained machine learning model from the MATLAB® workspace into Classification Learner and Regression Learner in one of two ways:

  • On the Learn tab, in the File section, select New Session > From Trained Model.

  • On the Learn tab, in the File section, click Import Model.

After you import a trained model into the app, you can:

  • Assess model performance using a test data set.

  • Explain model behavior using interpretability plots.

  • Export the model for deployment.

For a list of supported models and restrictions, see Import Trained Model from Workspace into Classification Learner or Regression Learner.

Machine Learning Apps: Train customizable neural networks (requires Deep Learning Toolbox)

Create a customizable neural network model in Classification Learner or Regression Learner by selecting the model preset Fully Connected Customizable Neural Network or Residual Customizable Neural Network in the Models section of the Learn tab. To customize the model's neural network architecture, select the model in the models pane, and click Customize Network on the model Summary tab. The default model for the fully connected customizable network contains five fully connected layers (excluding the final fully connected layer for prediction). The default model for the residual customizable network contains one residual connection. For more information, see Customizable Neural Network Models and Edit Customizable Neural Network Using Network Editor in Classification Learner or Regression Learner.

Machine Learning Apps: Specify validation and test partitions at MATLAB command line

When you launch Classification Learner or Regression Learner from the MATLAB command line, use the ValidationPartition name-value argument to specify a cvpartition object that defines the validation scheme and the indexing for the validation sets. To define the indexing for the test data set, specify a cvpartition object using the TestPartition name-value argument. For more information, see Classification Learner and Regression Learner.

Machine Learning Apps: Export partitions and data sets

Export the partitions used to compute validation and test metrics, as well as the data sets, from the current session to the MATLAB workspace. In the Export section of the Learn tab, select Export > Export Partitions and Data Sets. For more information, see Export Partitions and Data Sets from Classification Learner or Regression Learner.

Machine Learning Apps: Export a feature ranking plot

Export a feature ranking plot or its data by first creating the plot using the Feature Selection button in the Options section of the Learn tab. Select Export Plot to Figure or Export Plot Data in the Export section.

 Functionality being removed or changed

Machine Learning Apps: Automatically select the number of predictors to sample in ensemble tree models

Behavior change

In Classification Learner and Regression Learner, when you create a draft tree model from the Ensemble Classifiers or Ensembles of Trees section of the Models gallery, the default setting for the number of predictors to sample is now Auto. For Boosted Tree and RUSBoosted Tree models, Auto is equivalent to the default Select All setting in previous releases. For Bagged Tree models, Auto sets the number of predictors to sample as the square root of the number of predictors in Classification Learner, and one third of the number of predictors in Regression Learner.

Machine Learning

 Build, share, and deploy machine learning workflows using machine learning pipelines

A machine learning pipeline is a set of connected steps, called components, that are executed in a specified order to process data and perform machine learning. You can use pipelines to develop, share, and deploy end-to-end machine learning workflows.

Use the built-in components for data processing, feature selection, and supervised learning, or convert your custom functions into pipeline components. You can combine components to create pipelines for various machine learning applications.

For an example of a pipelines workflow, see Create Simple Classification Pipeline. For more information about the available components and functionality, see Machine Learning Pipelines.

Machine learning pipeline for SVM classification. The pipeline consists of components for removing observations, normalization, one hot encoding, PCA, and SVM classification.

UMAP: Reduce and visualize high-dimensional data

Using the umap function, you can reduce high-dimensional data to a low-dimensional embedding in order to view natural clustering and perform exploratory data analysis. For details, see the function reference page.

Evaluate model performance on slices of data

Evaluate the performance of a model on subsets of data (data slices) by using the sliceMetrics function. The function creates a sliceMetrics object, which you can use to compute metrics (such as accuracy or mean squared error) on the data slices and their complements. Use the report object function to summarize the metrics in a table, and use the plot object function to visualize the metrics as bar graphs.

Synthetic Data Generation: Generate synthetic data for imbalanced data sets using SMOTE

You can use the synthetic minority oversampling technique (SMOTE) algorithm to generate synthetic data for binary classification. Using SMOTE can be helpful when you have imbalanced data, that is, when one class contains many more observations than the other.

Regardless of the method used (SMOTE or binning), the synthesizeTabularData functions can return synthetic observations for one or two classes.

XGBoost Importer: Perform inference using imported XGBoost models

Import pretrained XGBoost regression and classification models into MATLAB and perform inference using the new importModelFromXGBoost function.

Incremental Learning: Create a model for incremental normalization

The incrementalNormalizer function creates a model object that is suitable for incremental normalization. Unlike the zscore function, for which you must provide all of the data before computing z-scores, incrementalNormalizer allows you to update the weighted predictor mean and standard deviation estimates incrementally and return z-scores by supplying chunks of data to the incremental fit function.

You can create the following types of incremental normalization model objects:

After creating a model object, you can train the model and calculate z-scores in real time as the model accesses data, either per individual observation or specified batch size.

  • The fit function updates the model object with information computed from the input model and data. The function optionally returns the z-scores.

  • The transform function transforms the input data into z-scores by using the incremental normalizer model.

  • The reset function resets all learned parameters of the model.

Generate counterfactual examples to better understand binary classifier decisions

Gain insight into binary classifier decisions by generating counterfactual examples using the counterfactuals function. Counterfactual examples identify the minimal modifications needed to change the predicted label of a given observation.

 Quantile Regression: Perform hyperparameter optimization with multiple quantiles

You can optimize the hyperparameters of a quantile regression model with multiple quantiles. Specify the OptimizeHyperparameters and Quantiles name-value arguments in the call to fitrqlinear or fitrqnet.

 Compatibility Considerations

If you perform Bayesian hyperparameter optimization when creating a RegressionQuantileLinear or RegressionQuantileNeuralNetwork object, the HyperparameterOptimizationResults property contains a SupervisedLearningBayesianOptimization object. In previous releases, the stored object is a BayesianOptimization object.

Hyperparameter Optimization: Perform cost-sensitive hyperparameter optimization for classification models

You can optimize classification model hyperparameters with respect to misclassification cost. When you use a classification fit function, specify the OptimizeHyperparameters and HyperparameterOptimizationOptions name-value arguments. In the HyperparameterOptimizationOptions structure or object, set the LossFun value to "classifcost", "mincost", or "auto-cost", depending on the classification fit function.

Depending on the type of supervised learning fit function you use to perform hyperparameter optimization, you can set the LossFun value to "auto-cost", "classifcost", "classiferror", "mincost", "mse", or "quantile".

GPU Support: Specify GPU arrays for cvpartition (requires Parallel Computing Toolbox)

You can now partition data for cross-validation on a GPU by supplying GPU array stratification variables or custom test sets to the cvpartition function. The object functions repartition, summary, test, and training accept the resulting cvpartition object and can execute on a GPU.

For a full list of Statistics and Machine Learning Toolbox™ functions that accept GPU arrays, see Function List (GPU Arrays).

GPU Support: Specify GPU arrays for Gaussian process regression models (requires Parallel Computing Toolbox)

The following functions now support GPU arrays, enabling you to execute the functions on a GPU:

For a full list of Statistics and Machine Learning Toolbox functions that accept GPU arrays, see Function List (GPU Arrays).

GPU Support: fitcecoc supports kernel learners for gpuArray inputs (requires Parallel Computing Toolbox)

You can specify kernel learners when you create a CompactClassificationECOC or ClassificationPartitionedKernelECOC model object by passing gpuArray data to fitcecoc.

To view the limitations of this support, see the Extended Capabilities sections in the documentation for fitcecoc, CompactClassificationECOC, and ClassificationPartitionedKernelECOC.

GPU Support: Specify GPU arrays when computing Shapley values using the Linear SHAP algorithm (requires Parallel Computing Toolbox)

The shapley and fit functions accept GPU array input arguments when the machine learning model is a regression or binary classification linear model listed below, and the function uses an interventional algorithm (Method="interventional").

The supported models are:

Performing Shapley value computations on a GPU is typically faster than on a CPU when you have a large number of query points and you do not specify an output function (OutputFcn=[]).

Neural Networks: Specify custom neural network architecture for data with categorical predictors

Specify a custom neural network architecture for categorical predictors using the Network argument of the fitcnet and fitrnet functions. Specify the neural network architecture as an array of deep learning layers or as a dlnetwork object. Use this argument when the fitcnet and fitrnet functions do not provide the neural network architecture that you need for your task such as neural networks with skip-connections.

Neural Networks: Convergence information and training history for custom neural network architectures

For ClassificationNeuralNetwork and RegressionNeuralNetwork objects fit using a dlnetwork or layer array that specifies the neural network architecture:

  • The ConvergenceInfo property now contains values for the ValidationChecks field.

  • The TrainingHistory property now contains the ValidationChecks and Time variables.

fitrnet Function: Use validation data when the regression model has multiple response variables

When creating a neural network regression model with multiple response variables, you can use validation data by specifying the ValidationData value in the call to fitrnet. For further customization, you can also specify the ValidationFrequency and ValidationPatience name-value arguments.

Categorical Predictors: Count the number of predictors in tabular data after encoding the categorical variables

Count the number of predictors in tabular data after encoding the categorical variables using the countPredictorsAfterCategoricalEncoding function.

You can use this function to help define custom neural network architectures for the fitcnet and fitrnet functions.

sequentialfs Function: Compute the criterion value for each candidate feature set in parallel (requires Parallel Computing Toolbox)

Compute the criterion value for each candidate feature set in parallel when no cross-validation is performed. In previous releases, sequentialfs performs only cross-validation in parallel. That is, the function runs computations in parallel when both of the following are true:

  • The UseParallel field of the options structure (Options) is set to true.

  • The cross-validation option (CV) uses more than one test set, or the number of Monte Carlo repetitions (MCReps) is greater than 1.

Starting with this release, provided that UseParallel=true in the options structure, the function computes the criterion value for each candidate feature set in parallel when no cross-validation is requested or when cross-validation is requested with one Monte Carlo repetition.

Example of handling class imbalance in binary classification

A new example, Handle Class Imbalance in Binary Classification, shows how to handle class imbalance in binary classification using decision thresholding, random undersampling, random oversampling, and SMOTE (Synthetic Minority Oversampling Technique).

 Functionality being removed or changed

Hyperparameter Optimization: Store Bayesian optimization results in a new object when using a supervised learning fit function

Behavior change

If you perform Bayesian hyperparameter optimization by using a supervised learning fit function, the optimization results are stored in a SupervisedLearningBayesianOptimization object. In previous releases, the optimization results are stored in a BayesianOptimization object.

A cross-validated neural network classification model is a ClassificationPartitionedNeuralNetwork object

Behavior change

A cross-validated neural network classification model is a ClassificationPartitionedNeuralNetwork object. In previous releases, a cross-validated neural network classification model is a ClassificationPartitionedModel object.

You can create a ClassificationPartitionedNeuralNetwork object in two ways:

  • Create a cross-validated model from a neural network classification model ClassificationNeuralNetwork by using the crossval function.

  • Create a cross-validated model by using the fitcnet function and specifying one of the name-value arguments CrossVal, CVPartition, Holdout, KFold, or Leaveout.

A cross-validated quantile neural network regression model is a RegressionPartitionedQuantileNeuralNetwork object

Behavior change

A cross-validated quantile neural network regression model is a RegressionPartitionedQuantileNeuralNetwork object. In previous releases, a cross-validated quantile neural network regression model is a RegressionPartitionedQuantileModel object.

You can create a RegressionPartitionedQuantileNeuralNetwork object in two ways:

  • Create a cross-validated model from a quantile neural network regression model RegressionQuantileNeuralNetwork by using the crossval function.

  • Create a cross-validated model by using the fitrqnet function and specifying one of the name-value arguments CrossVal, CVPartition, Holdout, KFold, or Leaveout.

rocmetrics and perfcurve Functions: Compute performance curves when a positive class is missing

Behavior change

You can now compute performance curves using rocmetrics or perfcurve when a positive class is missing from the true class labels. For example, each function can return metrics when a particular class appears in the training data but not in the validation or test data. Some returned metrics might have NaN values.

Statistics

Design of Experiments: Create a fractionalFactorialDOE object to generate an experiment design

Generate a two-level fractional factorial experiment design by using the fractionalFactorialDOE function to create an object of the same name. The object properties include information about the design, model, and factors used to generate the design. After creating the object, you can use the fitlm function to fit a linear model to the design points. Use the fractionalFactorialTypes function to return a table containing the resolution level and maximum number of runs for all possible two-level fractional factorial design types for both a set of factors and an experiment model.

Analysis of Lifetime Data: Create an accelerated life testing model

Create an AcceleratedLifeModel model object for accelerated life testing by using the fitacclife function. The function fits an accelerated life model to input data that contains stressor levels and their corresponding failure times. After creating the object, you can use it to generate plots and compute mean failure times, distribution functions, and failure time probabilities at specific stressor levels.

  • Compute distribution functions using the distfcn and icdf functions, and create distribution plots with the distplot function.

  • Compute and plot mean failure times at stressor levels using the meanfailtime and meanfailplot functions.

  • Plot predicted failure time probabilities at stressor levels using the probplot function.

  • Compute failure time acceleration factors at stressor levels relative to a baseline stressor level using the accelfactor function.

  • Compute confidence intervals for the fitted model coefficients using the coefci function.

Linear Regression: Find a minimum or maximum response value and calculate predictor values that yield a specified response (requires Optimization Toolbox)

Two new object functions are available for LinearModel and CompactLinearModel objects:

  • optimizeResponse: Use this function to find a minimum or maximum response value for a linear regression model and the predictor values for that response value (additionally requires Global Optimization Toolbox if the model includes interaction terms with categorical predictors whose values are not fixed using CategoricalValues).

  • matchResponse: Use this function to calculate the predictor values that correspond to a specified response value and have the smallest response variance (additionally requires Global Optimization Toolbox if the model includes categorical predictors whose values are not fixed using CategoricalValues).

GPU Support: Specify GPU arrays for noncentral chi-square distribution (requires Parallel Computing Toolbox)

The ncx2cdf, ncx2pdf, ncx2inv, and ncx2stat functions now support GPU arrays, enabling you to execute them on a GPU.

For a full list of Statistics and Machine Learning Toolbox functions that accept GPU arrays, see Function List (GPU Arrays).

GPU Support: Specify GPU arrays for Rician distribution (requires Parallel Computing Toolbox)

The fitdist and mle functions now support GPU arrays for Rician distributions.

For a full list of Statistics and Machine Learning Toolbox functions that accept GPU arrays, see Function List (GPU Arrays).

mhsample Function: For improved performance, explicitly specify Gaussian or Student's t proposal distribution

The Proposal name-value argument enables you to explicitly specify the distributions in this table by their name. You can tune the sampler by adjusting corresponding distribution parameters using name-value argument syntax.

NameProposal DistributionParameters
"Gaussian"Multivariate Gaussian distribution

  • Center — Optional mean vector

  • Scale — Required covariance matrix

"t"Multivariate Student's t distribution

  • Center — Optional center vector

  • Scale — Required scale matrix

  • DegreesOfFreedom — Optional degrees of freedom parameter

When you specify these distributions by setting Proposal instead of setting the PropPDF or LogPropPDF to a custom function handle to either of these distributions, mhsample might show improved performance. For example, the code below compares the run-time performance of mhsample when implicitly specifying a scale 1 Gaussian proposal distribution by using a custom function and when explicitly specifying a scale 1 Gaussian proposal by using the Proposal and Scale name-value arguments. The mhsample function is 20x faster when a Gaussian proposal is explicitly specified. 

F  = @(x) -x.^2 + log(2 + sin(5*x) + sin(2*x)); % Target stationary distribution log-PDF
numSamples = 500000;                            % Length of the MCMC chain
propPDF = @(x,y)mvnpdf(x,y,1);                  % Proposal probability density function
propRNG = @(x)mvnrnd(x,1);                      % Proposal random number generator

% Implicitly specify Gaussian proposal by specifying a function handle 
% to Gaussian random number generator
rng(1,"twister")
timeitimplicit = @()mhsample(0,numSamples,LogPDF=F,PropPDF=propPDF, ...
    PropRND=propRNG,Symmetric=true);
timecgp = timeit(timeitimplicit);

% Explicitly specify Gaussian proposal, under the same conditions, by naming 
% the proposal distribution using Proposal.
rng(1,"twister")
timeitexplicit = @()mhsample(0,numSamples,LogPDF=F,Proposal="Gaussian", ...
    Scale=1);
timegp = timeit(timeitexplicit);
The code was timed on a Windows 11 Enterprise, Intel(R) Xeon(R) W-2133 CPU @ 3.60GHz 3.60 GHz system. Experienced efficiency is problem dependent.

 Functionality being removed or changed

mhsample Function: NumChains name-value argument replaces nchains

Behavior change

When you call mhsample, to specify the number of Markov chains to generate from the Metropolis-Hasting algorithm, use the NumChains name-value argument instead of nchains. Although you should update your code to use NumChains instead of nchains, there are no plans to remove nchains.

Visualization

controlchart Function: Create P', NP', U', and C' charts

The controlchart function can now plot Laney P', NP', U', and C' charts, which adjust the control limits for subgroup sizes and inter-subgroup variations.

confusionchart Function: Customize confusion chart display

You can now customize the rotation of row (y-axis) and column (x-axis) labels of ConfusionMatrixChart objects:

You can also specify to display zero values in the chart using the ZerosVisible property.

To create a confusion chart, use the confusionchart function. For more information about how to set these properties, see ConfusionMatrixChart Properties.

 Functionality being removed or changed

controlchart Function: Set the center line value when specifying Limits and Rules

Behavior change

When you specify both the Limits and Rules name-value arguments for controlchart, the center line value is equal to the middle element value of Limits. In previous releases, the center line value is equal to the arithmetic mean of all measurement values.