Tech giant Google now aspires to put the power of Machine Learning in the hands of every enthusiast learner irrespective of the fact whether they hold a fancy degree or not.
Earlier in the month, Google published a blogpost talking about “AutoML” for “auto-machine learning,” a reinforcement learning approach that provides one A.I. the capability of becoming the architect of another one, and direct its own development without having to depend on a human engineer for input.
The approach basically involves a controller neural net proposing a “child” model architecture, which can then be trained and evaluated for quality on a particular task. The feedback received is then used to train the controller on how it can improve its proposals for the next round. This whole process is repeated thousands of times, which eventually leads to the controller learning to assigning high probability to areas of architecture space that achieve better accuracy on a held-out validation dataset, and low probability to areas of architecture space that score poorly.
While AutoML might sound like something that would lead to evolution of the singularity, but that’s not what Google aims to do. With AutoML, Google plans to expand the number of developers able to make use of machine learning by reducing the expertise needed.
According to a statement by Google CEO Sundar Pichai, “We hope AutoML will take an ability that a few Ph.D.s have today and will make it possible in three to five years for hundreds of thousands of developers to design new neural nets for their particular needs.”
For the uninitiated, a neural net is basically a computer system modeled after the human brain.
Usually, in order to solve a problem with machine learning, a human engineer is required to provide a starting neural network that is already structured to do the basic type of computation the problem requires. However, AutoML starts out by trying a number of possibly suitable algorithms, essentially testing totally different neural network architectures, and then scores each against the goal. With no human intervention, the process could end up zeroing in on both the best mathematical approach to the problem and the best implementation of that mathematical approach in a short period of time. In fact, the final neural net doesn’t even need to make use of just one of these algorithms either, and can include individual elements of multiple, if that turns out to be more useful.
In an email interview to MIT Technology Review, Pichai said that he his confident that AutoML would fit in with Google’s strategy of positioning its cloud computing services as the best place to build and host with machine learning. Google is meticulously working on earning new customers in the corporate cloud computing market, where it currently lags behind market leader Amazon and second-place Microsoft.
Several experts might argue that Google’s vision of making use of neural networks to design other neural networks sounds quite familiar, considering the fact that making programs to edit the code of other programs is the precise definition of machine learning. But, what makes AutoML stand out from the rest is how early in the process of designing a neural net it begins to intervene. In addition to this, not only will AutoML refine the simple models that already exist, but it will also select those models, and then refine them on its own. So, we can say, that AutoML could turn out to be a more full-featured version of what a normal machine learning was always supposed to be but never did.
Though in theory, it might seem that the AutoML approach should be able to design more efficient neural nets, but in being more efficient those Artificial Intelligence creations, it could also end up being more difficult for humans to understand.
In fact, in its blogpost, Google even demonstrated the human attempt at the best, most efficient neural network to tackle a particular database of images and compared it with AutoML’s neural network, featuring extra nodes that Google says resemble improvements recently proposed by human researchers.
The images above demonstrate that when given a large database of images to categorize, AutoML designed a neural net that was quite similar but slightly superior to the one designed by Google’s human engineers. What’s really interesting about this sort of design by proxy is that the engineers looking at AutoML’s neural network didn’t actually know the differences between it and their own were really improvements; since they hadn’t come up with the neural network themselves, they weren’t totally sure at first.
It is important to understand here that that AutoML doesn’t aim to drive humans out of the development process but sheerly wants to let A.I. continue revolutionising the world at the same pace at which we have been witnessing for years now. Unfortunately, difficulty of coding neural networks is slowly becoming a big problem for an industry that depends heavily on talent. AutoML aims to lower the entry bar level for the coming generation of machine learning students, at least when it comes to the most common and simplest of the applications.