AutoKeras: The Killer of Google’s AutoML

comments

Google AI has finally released the beta version of AutoML, a service that some are saying will change the way we do deep learning entirely.

Google’s AutoML is a new cloud software suite of Machine Learning tools. It’s based on Google’s state-of-the-art research in image recognition called Neural Architecture Search (NAS). NAS is basically an algorithm that, given your specific dataset, searches for the most optimal neural network to perform a certain task on that dataset. AutoML is then a suite of machine learning tools that will allow one to easily train high-performance deep networks, without requiring the user to have any knowledge of deep learning or AI; all you need is labelled data! Google will use NAS to then find the best network for your specific dataset and task. They’ve already shown how their methods can achieve performance that is far better than that of hand-designed networks.

AutoML totally changes the whole machine learning game because for many applications, specialized skills and knowledge won’t be required. Many companies only need deep networks to do simpler tasks, such as image classification. At that point they don’t need to hire 5 machine learning PhDs; they just need someone who can handle moving around and organising their data.

So is AutoML a beautiful silver bullet that let’s any company or individual easily do AI?!

…. Not so fast.

To use Google’s AutoML for computer vision, it will cost you USD $20 per hour. That’s crazy! You won’t even be sure that you’ll get much better accuracy than your own hand-designed network until you pay and try it out. It’s interesting to see that Google is monetizing this when in the past, both with Google and in general the AI community, people have favored open source to be able to share the knowledge with everyone.

And that’s exactly where Google’s AutoML will lose: open source.

Enter AutoKeras, an open source python package written in the very easy to use deep learning library Keras. AutoKeras uses ENAS, an efficient and most recent version of Neural Architecture Search. You can quickly and easily install the package with a pip install autokeras and voila, you’re ready to do your own architecture search on your own dataset … for free.

Since all of the code is open source, you can even play around with some of the parameters if you want to do something really custom. It’s all Keras so the code is easy to understand and dive into, allowing developers to quickly create accurate models and researchers to take a deeper dive into architecture search.

AutoKeras has everything a great open source project should have: quick install, easy to run, lots of examples, easy to modify, and you even get to see the network model that NAS found out at the end! If you prefer a TensorFlow or Pytorch, those are also available here and here!

I’d definitely recommend trying out AutoKeras or any of the other implementations as a ridiculously cheaper alternative to AutoML. Perhaps Google is doing something more under the hood for AutoML over the open sourced stuff, but given the very small performance difference between NAS models and hand designed ones I doubt that one would get substantial enough gains to be worth paying such a steep price.

Deep learning and AI in general is such a powerful technology that we shouldn’t be trying to hide it behind such high paywalls. Yes Google, Amazon, Apple, Facebook, and Microsoft are all businesses that need to make money to survive and compete. But something like this, where the research paper is public and we have deep learning libraries available to quickly replicate the methods, it simply doesn’t make sense to try and block it from people when it can so easily be made open.

There’s also a potentially bigger problem here: that knowledge itself is being hidden. One of the great things about the recent trend in AI has been the decision of much of the research community to quickly post their work openly on sites such as Arxiv to share with the community and get feedback. Even more so, there has been a growing trend to post the research code on Github for reproduction and in general further usage of the algorithms in research and real-world applications. Yet here we still see such research being put behind a paywall.

Sharing science helps with progress and increasing the general knowledge of everyone. One thing’s for sure: knowledge should be open source. It’s better for everyone.

Like to read about tech? Follow me on Twitter where I post all about the latest and greatest tech!

 
Bio: George Seif is a Certified Nerd and AI / Machine Learning Engineer.

Original. Reposted with permission.

Related:

• The Current State of Automated Machine Learning
• Selecting the Best Machine Learning Algorithm for Your Regression Problem
• The Keras 4 Step Workflow