- 16th April 2018
- Posted by: Manolis
An AI that can code its own ‘Child AI’
If you have been following the tech for a while it would be obvious to you that artificial intelligence has made some significant strides in the past couple of years. But today we’re going to look at something a little bit different.
Google has just created an AI that is capable of creating its own AI that performs better than anything else made before in its field. This sounds impossible but has just happened. Google calls it Automatic Machine Learning (Auto ML in short). It’s the work of team of researchers at Google brain. The researchers at Google used reinforcement learning to create AI inception as described by Google CEO Sundar Pichai :-
Today designing better machine learning models is really time consuming and it’s a painstaking effort of a few engineers and scientists mainly machine learning PhD’s. We want it to be possible for hundreds of thousands of developers to use machine learning. So what better way is to do this than getting neural nets to design better neural nets. We call this approach Auto ML, it’s learning to learn. We are already approaching state-of-the-art in standard tasks like so far image recognition. So whenever I spend time with the team and think about neural nets building their own neural nets, it reminds me of one of my favourite movies Inception and I tell them we must go deeper.
So instead of spending painstaking hours coding to create an Artificial Intelligence to do a certain task. Auto ML actually automates this process. Here is a quote from Google Research Team
In our approach (which we call ‘Auto ML’), a controller neural net can propose a ‘child’ model architecture, which can then be trained and evaluated for quality on a particular task. That feedback is then used to inform the controller how to improve its proposals for the next round. We repeat this process thousands of times generating new architectures, testing them, and giving that feedback to the controller to learn from.In our approach (which we call ‘Auto ML’), a controller neural net can propose a ‘child’ model architecture, which can then be trained and evaluated for quality on a particular task. That feedback is then used to inform the controller how to improve its proposals for the next round. We repeat this process thousands of times generating new architectures, testing them, and giving that feedback to the controller to learn from.
In a strange way you can think of it like a parent giving birth to a child. Then teaching that child how to perform a particular task through giving it tests and telling it when it passed or failed. The parent in this case is Auto ML and the child is another neural network called NASNet. It is capable of recognizing objects such as people, cars and most of this things in real time video. This automated AI outperformed any other computer vision that was made before it. NASNet can perform with 82.7% accuracy, 1.2% better than all other efforts while being 4% more efficient.
An AI built a child AI that performed better than any other artificial intelligence created by humans and it was more efficient. Google Quoted
We hope that the larger machine learning community will be able to build on these models to address multitudes of computer vision problems we have not yet imagined.
We think this can inspire new types of neural nets and make it possible for non-experts to create neural nets tailored to their particular needs, allowing machine learning to have a greater impact to everyone.
Applications of Auto ML ?
Obvious applications for such a system like Auto ML include self-driving vision support. Enabling object recognition for robots that work in close quarters with humans. This all has huge implications mainly because if an AI can create systems that are accurate while remaining efficient, it can really take the hard work out of the field of machine learning.