AutoML: Where Machine Learning is Headed Next

Machine learning is not a new concept in cutting edge tech society and is used in a wide variety of advanced tech applications, namely, targeted advertising and data management. Until recently, a popular shopping app in Japan, called Mercari, used machine learning in order to classify photographs. However, a new system, called automated machine learning, or AutoML, rendered the app’s original methods a thing of the past, achieving a whopping 15% increase in accuracy, and motivating Mercari to make the full switch over to AutoML

What Is Automated Machine Learning Exactly? 

Machine learning has made significant changes in fields such as healthcare, retail financial services and even transportation, commonly being used in research and development, as well as enterprising. However, traditional machine learning requires a significant amount of human power, leaving many businesses unable to make use of this powerful tool.

It’s here where AutoML shines, as AutoML automates the entire process it would take to apply machine learning to a problem or task, opening the door for smaller companies and even non-experts in the field. Traditional machine learning requires four steps from start to finish: reading and merging, preprocessing, optimization, and application. AutoML focuses mainly on data collection and prediction, effectively the first, and last step of traditional machine learning, as these are easily automated already. In doing this, AutoML delivers a model that’s already optimized and more accurate than older methods.


Is There a Need for AutoML?

As stated previously, many businesses struggle to supply the human power needed to apply traditional machine learning, despite the clear benefits it could provide them. Not only does machine learning require a team of experienced data scientists, but also requires a decision on what model of machine learning would work best for the business, which means that the data scientists, who already claim a premium salary, would need more experience with the business in order to be effective. The main benefit of AutoML is that it fills the demand for a system that can be used off the shelf by those with less experience while still offering the same boost to their business.


What Are the Advantages of AutoML?

There are three major advantages that AutoML has over traditional methods:


  • Increased Productivity:  By automating repetitive tasks, data scientists are able to focus more on the problem that the model is trying to solve, rather than the model itself.
  • Avoids Errors:  Manual data entry inevitably comes with manual errors. Automating this pipeline helps to avoid most of these problems.
  • Democratization:  In cutting down the required workforce required to use machine learning, AutoML makes the platform available to a much wider user base.



7 of the Most Popular AutoML Frameworks Available



  • MLBox or Machine Learning Box.  This is an automated Python library that uses machine learning to read and distribute data. The features of MLBox include the following.



  • Gives you a robust and large feature selection including accurate hyperparameter optimization and leak detection,



  • Can preprocess, clean, and format data via fast reading and distribution,



  • Has modern predictive models like LightGBM, Stacking, and Deep Learning for classification and regression,



  • Can provide predictions based on model interpretations.



Has been tested on Kaggle with a rank of 85/2488 which is excellent. MLBox comes with 3 sub-packages which includes pre-processing for reading and processing data, optimization for testing and cross-validating, and prediction capabilities. It can be installed on Linux at the moment but Windows and MacOS support will be included in the future.



  • Auto-Sklearn. This automated machine learning package frees up the user from having to make algorithm selections or do any type of hyperparameter tuning as it is built on top of Scikit-learn. Features of this package include numeric standardization, one-hot coding, and uses models for classification and regression problems. It works by creating a pipeline and using Bayesian hyperparameter optimization for meta-learning and automated ensemble construction for configurations. Unfortunately, it cannot be applied to deep learning systems on large datasets.  It only works with Linux machines at the moment. 
  • Tree-Based Pipeline Optimization Tool or TPOT for Short. This automated machine learning tool is Python based, uses and optimizes on machine learning pipelines that use genetic programming. It extends on the Scikit-learn framework but uses its own classifier and regressor methods. It works by exploring thousands of pipelines and uses the best one for data. Due to this, it cannot process natural language inputs automatically and it cannot process categorial strings as these must be integer-encoded first before being passed as data.
  • H2O. This is an open-source machine learning platform from that uses in-memory machine learning with support for R and Python. It has support for the most widely used machine learning and statistical algorithms, including deep learning, gradient boosted machines, and generalized linear models. The automatic machine learning module within H20 will use its own algorithms to build a pipeline after performing an exhaustive search of its own engineering methods and model hyperparameters. This is how it creates an optimized pipeline for your data.  This platform is popular because it can automate some of the most difficult machine learning workflows and data science, including model validation, model tuning, and model deployment as well as model selection.
  • AutoKeras. Another open-source automated machine learning library which is based on Keras deep learning framework by Data Lab. It gives users the ability to automatically search for hyperparameters and architecture for deep learning models. It is an easier and simpler platform to use as it is based on Scikit-learn API but simplifies the process through automated neural architecture search algorithms. It is compatible with Python.
  • Cloud AutoML. The Cloud AutoML is from Google and is for developers who have limited machine learning knowledge. It is to aid developers in creating high-quality models specific to business needs. It utilizes transfer learning and neural architecture search technology. It is simply to use as it has a graphical user interface that gives developers the ability to train, evaluate, improve, and deploy models based on unique data.  Unfortunately, this program is not open-source and the price point varies based on which package you choose.
  • TransmogrifAI.  This is another open-source automated machine learning library and is an end-to-end library for structured data that is written in Scala. This machine learning library powers Einstein, the flagship product of Salesforce and it runs on top of Apache Spark. It is extremely good at training machine models with minimal tuning and it can build modular machine learning workflows. It requires Java and Spark to run though.


What is the Future of AutoML?

Put simply, AutoML was built to automate repetitive tasks such as pipeline creation and hyperparameter tuning, just like robots in a factory, for the purpose of allowing data scientists to focus on the issue at hand and how the model is affecting it, rather than the model itself. An additional bonus to this is that the technology is available to a larger network of users, rather than just the businesses with enough resources to commit to it. Ultimately, AutoML will play a large part in the future of machine learning, especially if it continues to advance and achieve breakthroughs.


Comments are closed.