Google Built AI That Can Build Its Own AI

Google Built AI That Can Build Its Own AI

KitGuru Says: At this point, I'd question as to whether Google took direct inspiration from Terminator to create Skynet, I mean, NASNet from its concept, directly to its naming. "These tools automatically sort through a huge range of alternatives relevant to some machine learning task", explained James Kobielus, lead analyst for data science, deep learning and application development at SiliconANGLE Wikibon. AutoML acts as a controller neural network that can create a "child" network to execute a specific task. The evaluation is then fed back to the controller, which uses the information to propose a new model.

When it was unveiled in May, AutoML tried its luck on two well-known, but relatively small datasets: image recognition with CIFAR-10 and language modeling with Penn Treebank.

Observers note that building machine learning models remains a costly, time-consuming and computing intensive process. But perhaps the most intriguing applications have yet to be identified.

Recently, the Google Brain's team chose to throw a challenge to the AI AutoML of creating a "child" that outdid all of its human-made counterparts by using an approach called reinforcement learning.

Company investigators recently outlined an effort to scale AutoML from small neural networks to "larger, more challenging datasets", including ImageNet image classification and an object detection framework called COCO. It was also on par with the best ever reported ImageNet result, based on the SENet neural network.


According to studies, the Google "brain" can pick out these objects with a 82.7% accuracy rate. Moreover, it performs 1.2% better than all previously published results. Additionally, a less computationally demanding version of NASNet outperformed mobile platforms by 3.1%.

The researchers suspect that the features learned by NASNet on ImageNet and COCO can be reused in many other computer vision applications, some of which can be applied to real-world applications, such as self-driving cars and security monitoring.

"When we look at the progress that has been made over the previous year in AI, we think that Google has continued to distance itself from its competition", Windsor emphasized in a research note issued on Monday (Dec. 4).

The Google researchers utilised reinforcement learning to allow the parent AI to teach NASNet how to detect and identify objects with extreme accuracy, higher than that of human-made design.

Related:

Comments

Latest news

Mark Cavendish set to compete in Tour de Yorkshire
The Tour de Yorkshire is a legacy from that Grand Depart and was a three-day race for its first three incarnations from 2015-17. Yorkshire is due to host the 2019 World Road Championships, when Harrogate will serve as the main hub for race finishes.

Pigeons Have Abstract Intelligence and Can Understand Concepts of Space and Time
How could that be tested? "Basically, looking for neurons that seem to be encoding both the length and duration of the line". But this only gets you so far, so the researchers added in other lengths and different time periods to mix things up a bit.

St Paul's security alert: Area on lockdown after suspicious package found
She said: 'There are dogs and police everywhere and all around St Paul's station has been cornered off. St Paul's Tube station was also closed while police looked into the incident. "More when we have it".

Seven killed in Syria bus bombing
A bomb blast killed eight people and injured 16 others on a bus in Syria's government-held Homs city on Tuesday, state media said.

Volkwagen goes rallying with the new Polo GTI R5
VW is no stranger to customer competition, with the company selling the Golf GTI TCR touring vehicle to teams around the world. This enables the 1,320 kg (2,910 lbs) rally vehicle to rocket from 0-100 km/h (0-62 mph) in just 4.1 seconds.

Other news