What Is AutoML and Why Is It Important?

Automl Featured

Automated machine learning (AutoML) is a whole new approach in machine learning where “parent” neural networks design “child” neural networks, taking humans out of the picture.

Using AutoML, even laypersons can train AI systems to perform any tasks without having to know machine learning.

Presently, Google, Microsoft and Amazon are trying to monetize AutoML with their B2B clients, but all the core research is available online for free. There are plenty of open-source frameworks that individuals can use in their machine learning applications.

History of AutoML

The concept of AutoML was first proposed in 2014 at a little-known workshop in the University of Freiburg, Germany. The organizers came with the idea of developing off-the-shelf machine learning methods that would not require any human experts.

Automl Workshop 2014 Opt

In May 2017 Google adopted the term AutoML to describe a machine learning model with self-replicating capabilities. Google CEO Sundar Pichai’s vision was that ordinary developers should be able to design AI systems using AutoML, without wasting time or growing new skills.

The concept soon spread like wildfire, and we now have several companies developing their own AutoML frameworks.

Basic Concepts in AutoML

For automated tools to compete with machine learning experts, they should be able to replicate the algorithms and optimize hyper-parameters such as learning rates.

According to the original AutoML website at the University of Freiburg, the following toolkits help teach machines the intelligence to design their own versions.

  • Sequential Model-Based Algorithm (SMAC): it uses predictive models to determine the important input variables. This helps in optimizing algorithm parameters.
  • Auto-Sklearn: an out-of-the-box learning tool, Auto-Sklearn searches for the best machine learning datasets and utilizes past knowledge.
  • Bayseian Optimization and Hyperband (BOHB): a deep learning tool, BOHB uses Bayseian techniques and bandit-based methods to optimize the hyper-parameters.

All three toolkits are available as GitHub repositories at AutoML websited. They can run from Python and Java and come with several features for data mining and data analysis.

Different Versions of AutoML

As discussed before, several companies have their own AutoML frameworks for commercial use. Rachel Thomas, a co-founder at Fast.ai, calls them “drag-and-drop” machine learning interfaces.

Tweet Drag And Drop Machine Interfaces

Google has one such drag-and-drop interface which they call Cloud AutoML. It solves various problems such as translation between languages and video intelligence. Google is currently using the former in its online translation service and the latter in content discovery for YouTube. Clearly, AutoML is very important to its platform offerings.

Cloud Automl

Moreover, Google’s AutoML offering is only available to businesses and you have to pay $20 per hour to avail these services. So, if you’re looking for AutoML to integrate with your machine learning project, don’t fall for Google’s marketing gimmicks. There are open-source alternatives out there.

Auto-Keras is one good example of a complete AutoML software library compatible with Python 3.6. As per the website, you can use a Pip install command in Python to install and manage this software package.

Refer to the documentation on above website to familiarize yourself with essential commands to run on Python. It also has the link to a Gitter community where you can find further support.

Gitter Autokeras Opt

Applications of AutoML

AutoML finds plenty of use in IoT because it reduces the time to build a model.

  • In smart transportation, AutoML can eliminate human involvement in traffic analysis, parking and self-driving cars.
  • In healthcare and life sciences, AutoML can tackle the problems of patient management and design better drugs.
  • Gamers can use AutoML in the background to gain insights and predictions about their best moves.
  • Financial services can use AutoML to stay competitive because of its high-quality analytics.

In Summary

AutoML is a game-changer in IoT because of how easily it can replace human involvement in ML projects. Clearly, it looks like a classic case of AI replacing humans. However, you will always need a developer at the forefront, and AutoML will help her or him relax while the program self-replicates.

Resources: to know more about AutoML, you should check out a few important research papers at the University of Freiburg website.

What are your views on automating machine intelligence? Do let us know in the comments.

4 comments

  1. “you will always need a developer at the forefront”
    I wouldn’t bet on that. Human race will go to its grave believing that it is indispensable. What is to prevent AI from taking over the developer job? After all, AI will be able to do it better and faster. Currently humans are needed to create new copies of machines. Once machines learn how to replicate themselves, humans will become superfluous. I wonder how long it will be before machines start considering humans as a “nuisance”? And how long before humans become an “infestation”? Resistance is Futile!

    1. AutoML is a pretty innocent technology although I may have suggested otherwise as a tongue in cheek reference. Right now It is only used to take care of repetitive tasks. Imagine an automatic spot welder in an assembly line. Only difference, AutoML is more intelligent because it can identify correct input variables and algorithms for a given task or problem.

      But who does the initial programming? It’s a human being. Who designs the original spot welder? Without a human being behind the scenes, AutoML will always remain dumb and useless.

      So, the watchword is “repetition.” AutoML helps you avoid repetitive, low value decision making activities. That’s what I meant in regards to “taking humans out of the picture.”

      1. IF things continue as they are today, then you are right. But things never remain static. Engineers, scientists, programmers and others are always working on making improvements, if only to see if it can be done and how far they can push the envelope. They are forever competing to see who can build a better mouse trap. Just look around you. What are we using today that is at the same level of development as it was even five years ago?

        The first computers were nothing more than calculating machines to solve ballistics equations. Today we have IoT being incorporated into every device. We have AutoML. We have AI medical diagnostic machines/robots/programs, albeit crude in comparison to human doctors but nevertheless capable of making diagnosis.

        “But who does the initial programming?”
        The parents. But eventually the kids leave the parents and start learning on their own, changing their own “programming”. Humans will program AI initially but at some point AI will start programming itself.

        “Without a human being behind the scenes, AutoML will always remain dumb and useless.”
        You are somehow forgetting or overlooking what “ML” stands for – machine LEARNING. Once the learning process starts, it is like a snowball rolling down hill. It gets bigger and rolls faster as it goes along. Just look at the history of human learning. Today, a high school freshman knows more than Archimedes, Pythagoras and DaVinci combined. Like a snow ball rolling down hill, Learning cannot be stopped. The Church tried very hard. They excommunicated and burned Bruno. They made Galileo recant who, nevertheless, is reputedly said “E pur si muove”. The Holy Mother Church, in spite of all its edicts, excommunications and preaching could not prevent Copernicus from stopping the Sun while getting the Earth moving.

        The only way for AutoML to remain “dumb and useless” is to never let it learn too much and become more than industrial robots. The march of progress can be slowed but it cannot be stopped because there will always be somebody who will want to see “what will happen if”. Most, if not all, inventions were made to make human life easier, more convenient, to make whatever process more efficient. How efficient is human supervision of machines? Humans get tired, bored, their attention wanders, they need bathroom breaks. AI does not suffer from such shortcomings. AI supervising machines is much more efficient. Sentient, self-maintaining, self-programming AI/robots (sounds to me like a plastic and metal version of a human) are inevitable.

        How do you expect to have autonomous vehicles if a human must program every possible situation into them? The AI controlling autonomous vehicles must be able to learn the way human drivers learn, by experience and on the fly. Otherwise, when a new situation occurs, the AI will not know what to do and it will stop (at least I hope “stop” is the default condition) and wait for a human to program in the proper reaction for the condition that just occurred. Defeats the whole purpose of an autonomous vehicle.

        Neither human drivers not AI can be pre-programmed to handle all possible situations.

  2. “You are somehow forgetting or overlooking what “ML” stands for – machine LEARNING. Once the learning process starts, it is like a snowball rolling down hill. It gets bigger and rolls faster as it goes along. Just look at the history of human learning. Today, a high school freshman knows more than Archimedes, Pythagoras and DaVinci combined. Like a snow ball rolling down hill, Learning cannot be stopped. The Church tried very hard. They excommunicated and burned Bruno. They made Galileo recant who, nevertheless, is reputedly said “E pur si muove”. The Holy Mother Church, in spite of all its edicts, excommunications and preaching could not prevent Copernicus from stopping the Sun while getting the Earth moving.

    The only way for AutoML to remain “dumb and useless” is to never let it learn too much and become more than industrial robots. The march of progress can be slowed but it cannot be stopped because there will always be somebody who will want to see “what will happen if”. ”

    Didn’t the Late Vatican exonerate Bruno with Pope John Paul II issuing a belated apology on behalf of those 16th century Inquisitors? Not saying that free thinkers are having it easy in the 21st century either. Just take a look at what’s currently happening to Julian Assange. In a non-reactionary world, he would have been winning a Nobel prize. Unfortunately, the misguided rabble still calls the shots about what should happen to whom while the aristocracy gives a wink of approval.

    Coming to the point – I think you’re giving way too much importance to AutoML. You have nothing to fear of it anymore than a block of Python code. In fact, these are literally solid blocks of Python. If you check the GitHub repositories for any of AutoML’s components at the Auto-Keras website, you can see the code for yourself.

    You can run AutoML right now on your own PC – sorry I believe it must be Linux 🙂 But do give it a test drive.

    Let us not compare a block of Python code (which AutoML really is) with killer robots which is the threat you’ve spoken of. Mostly it is between the militaries.

    There is even a campaign to stop killer robots https://www.stopkillerrobots.org/learn/ which claims to have 25000 AI experts worldwide.

    Autonomous vehicles run on Convolutional neural networks (CNN) which are an extreme subset of deep neural networks. AutoML is small fry in comparison.

Comments are closed.