Kotlin for Machine Learning: Capabilities and Limitations

Kotlin for Machine Learning: A Modern Take on Data Science

The world of machine learning (ML) has long been associated exclusively with Python. However, with the growing popularity of Kotlin in Android development and backend, the community has begun actively adapting this language for Data Science tasks. Kotlin offers concise syntax, null-safety, and full compatibility with the Java ecosystem. In this article, we will explore how suitable Kotlin is for machine learning, what tools already exist, and what limitations a developer will have to face.

Why might Kotlin be interesting for ML?

Kotlin does not try to replace Python, but offers a number of advantages in certain scenarios:

• JVM Performance. Kotlin runs on the Java Virtual Machine (JVM), providing access to powerful libraries for multithreading and computation optimization.
• Conciseness and Safety. Less boilerplate code, strong typing, and the absence of NullPointerException make code more reliable.
• Java Ecosystem. You can use any Java libraries: Deeplearning4j, Apache Spark MLlib, ND4J, Weka, and others.
• Unified Stack. If your company writes backend in Kotlin, implementing ML modules will not require switching to another language.

Main Libraries for Machine Learning in Kotlin

Despite the youth of this direction, there are already several mature frameworks that cover the basic needs of an ML engineer.

KotlinDL — Native Library from JetBrains

KotlinDL (Kotlin Deep Learning) is an open-source library developed by JetBrains. It provides a high-level API for building neural networks, training, and inference. The library supports ONNX models, allowing you to import models from PyTorch or TensorFlow.

import org.jetbrains.kotlinx.dl.api.core.Sequentialimport org.jetbrains.kotlinx.dl.api.core.layer.Denseimport org.jetbrains.kotlinx.dl.api.core.activation.Activationsimport org.jetbrains.kotlinx.dl.api.core.optimizer.Adam

fun main() {    val model = Sequential.of(        Dense(128, activation = Activations.Relu),        Dense(64, activation = Activations.Relu),        Dense(10, activation = Activations.Softmax)    )

    model.compile(optimizer = Adam(), loss = "categorical_crossentropy")    println("Model created: ${model.summary()}")}

KotlinDL is well suited for image classification tasks, entry-level NLP, and experimenting with architectures.

Smile — Library for Classical ML

Smile (Statistical Machine Intelligence and Learning Engine) is a powerful Java library that is easily used from Kotlin. It includes algorithms: linear regression, SVM, random forest, k-means clustering, PCA, and much more.

import smile.classification.RandomForestimport smile.data.DataFrameimport smile.data.vector.DoubleVector

fun main() {    val x = arrayOf(        doubleArrayOf(1.0, 2.0),        doubleArrayOf(2.0, 3.0),        doubleArrayOf(3.0, 4.0)    )    val y = intArrayOf(0, 1, 0)

    val model = RandomForest.fit(x, y)    val prediction = model.predict(doubleArrayOf(2.5, 3.5))    println("Prediction: $prediction")}

Smile is ideal for tasks where deep learning is not required, but speed and reliability of classical algorithms are needed.

Multik — Multidimensional Arrays for Kotlin

Multik is an analogue of NumPy for Kotlin. It provides multidimensional arrays (tensors), linear algebra operations, statistical functions, and integration with OpenBLAS for accelerated computations.

import org.jetbrains.kotlinx.multik.api.*import org.jetbrains.kotlinx.multik.ndarray.data.D2Array

fun main() {    val a: D2Array<Double> = mk.ndarray(mk[1