This article was originally published at: https://www.blog.duomly.com/6-top-machine-learning-libraries-for-javascript-in-2019/

Usually, people apply machine learning (ML) methods and algorithms using one of two programming languages: Python or R. Books, courses, and tutorials about Machine Learning most often use one of these languages as well (or both).

Python is a general-purpose programming language used not only for Machine Learning but also for scientific computing, back-end Web development, desktop applications, etc. R is created primarily for statisticians. However, they have at least two common characteristics:

• They are suitable for non-programmers,
• They have comprehensive ML libraries.
• In many cases, ML algorithms are implemented in Fortran, C, C++, or Cython and called from Python or R.

Java is also used for Machine Learning, but usually by professional programmers.

During the last few years, JavaScript gained in popularity, and some very interesting Machine Learning libraries appeared enabling implementation of ML methods in browsers or on Node.js. Surprisingly, many of such libraries implement a lot of code in JavaScript itself.

This article presents several ML open-source libraries for JavaScript:

• ml.js,
• TensorFlow.js,
• brain.js,
• ConvNetJS,
• WebDNN,
• natural.

ml.js

ml.js is a comprehensive, general-purpose JavaScript ML library for browsers and Node.js. It offers the routines for:

• Bit operations on arrays, hash tables, sorting, random number generation, etc.,
• Linear algebra, array manipulation, optimization (the Levenberg-Marquardt method), statistics,
• Cross-validation,
• Supervised learning,
• Unsupervised learning.

Supported supervised learning methods are:

• Linear, polynomial, exponential, and power regression,
• K-nearest neighbors,
• Naive Bayes,
• Support vector machines,
• Decision trees and random forest,
• Feedforward neural networks, etc.

Besides, ml.js offers several unsupervised learning methods:

• Principal component analysis,
• Cluster analysis (k-means and hierarchical clustering),
• Self-organizing maps (Kohonen networks).

License: MIT.

TensorFlow.js

TensorFlow is one of the most popular Machine Learning libraries. It focuses on various types and structures of artificial neural networks, including deep networks, as well as the components of the networks. TensorFlow is created by Google Brain Team and written in C++ and Python. However, it can be used with several languages, including JavaScript.

TensorFlow is a very comprehensive library that still enables building and training models easily. It supports a huge variety of network layers, activation functions, optimizers, and other components. It has good performance and offers GPU support.

TensorFlow.js is a JavaScript ML library for use in browsers or on Node.js. It supports WebGL.

License: Apache 2.0.

brain.js

brain.js is a library written in JavaScript-focused on training and applying feedforward and recurrent neural networks. It also offers additional utilities, such as math routines necessary for neural networks.

It provides advanced options like:

• Using GPU to train networks
• Asynchronous training that can fit multiple networks in parallel
• Cross-validation that is a more sophisticated validation method

brain.js saves and loads models to/from JSON files.

License: MIT.

ConvNetJS

ConvNetJS is another library for neural networks and deep learning. It enables training neural networks in browsers. In addition to classification and regression problems, it has the reinforcement learning module (using Q-learning) that is still experimental. ConvNetJS provides support for convolutional neural networks that excel in image recognition.

In ConvNetJS, neural networks are lists of layers. It provides the following layers:

• Input (the first) layer
• Fully connected layer
• Convolution layer
• Pooling layer
• Local contrast normalization layer
• Classifiers loss (the output) layers: softmax and svm
• Regression loss (the output) layer that uses L2

It supports several important activation functions like:

• ReLU
• Sigmoid
• Hyperbolic tangent
• MaxOut

as well as the optimizers such as:

• Stochastic gradient descent
• Adadelta
• AdagradS
• ConvNetJS also provides a convenient way to save and load models to/from JSON files.

License: MIT.

WebDNN

WebDNN is a library focused on deep neural networks, including recurrent neural networks with LSTM architecture. It is written in TypeScript and Python and offers JavaScript and Python APIs.

It also provides the possibility of GPU execution in browsers.

A very convenient feature of WebDNN is the possibility to convert and use the models pretrained with PyTorch, TensorFlow, Keras, Caffemodel or Chainer.

License: MIT.

natural

natural is a JavaScript library for natural language processing used with Node.js.

It supports:

• Tokenization (breaking text into arrays of strings)
• Calculation of strings distances
• Matching similar strings
• Classification (naive Bayes, logistic regression, and maximum entropy)
• Sentiment analysis (currently in eight languages)
• Phonetic matching, inflectors, n-grams, etc.

License: MIT.

Conclusion

Both JavaScript and machine learning are gaining much attention and popularity during the last several years. Although initially created to enable dynamic behavior of Web pages, JavaScript becomes one of the languages of choice to implement and apply machine learning methods, especially in browsers or servers (Node.js).

This article provided the initial information on the availability of machine learning libraries for JavaScript.

Have a lot of fun exploring them!