Not logged in » Login

Recent Discussion:

GGiner | 16.09.2020, 19:02
Today I've spent several hours trying to access ServerView RAID Manager in one of our remote servers...
NunoCosta | 10.07.2020, 10:56
Hi Michel.Our troubleshooting team has no more information on this subject. However, if you need mor...
mlegrafleitas | 09.07.2020, 17:28
Hi Nuno,Very thanks for your response, the storage DX100 S5 is configure with 4 iSCSI CA for SAN red...
Oct 31 2018

Google Introduces AdaNet

Given the extensive media coverage of Artificial Intelligence (AI), Machine Learning (ML) and their miraculous capabilities that could help solve all earthly problems, it's somewhat surprising to find that the actual adoption rate isn't as big as you'd expect. The reason for that is pretty simple: Both AI and ML are fairly complicated technologies that typically require a lot of expert knowledge to be implemented correctly – knowledge most IT teams don't have at their command. Google's new framework AdaNet is supposed to solve this issue, or at least parts thereof.

Technically speaking, AdaNet is supposed to automate and thereby simplify the adoption of ML technologies and, by extension, typical AI implementations like neural networks. To achieve this, AdaNet employs an approach called "ensemble learning," which combines multiple ML algorithms in order to improve the reliability and quality of predictive insights generated by running data sets against these algorithms. But as one might expect, what sounds easy in theory is much harder to accomplish in real life. Or as Charles Weill, one of Google's AI gurus, put it in a blog post yesterday, it "requires its own level of domain expertise."

For those who are familiar with AI/ML and hope to apply such methods to help them make sense of the tons of data they obtain and generate on a daily basis, AdaNet provides a lightweight framework for automatically learning high-quality models with minimal intervention. The idea behind it was to enable ML practitioners to save the time and effort normally required to select an adequate neural network architecture and instead have an adaptive algorithm do the hard work of understanding interdependencies and arranging several subnetworks into a whole. AdaNet also is capable of adding subnetworks of different depths and widths to create a diverse ensemble, and trade off performance improvement with the number of parameters. What's more, it measures the ensemble loss for each candidate, and selects the best one to move onto the next iteration, that way ensuring the ensemble of subnetworks may branch out as a business requires. Finally, Weill asserts that AdaNet can provide so-called learning guarantees, i.e. assure that the ML mechanisms will undergo any training successfully. For best results, it should be combined with one of Google's specialized machine learning offerings, such as AutoML Vision (for image/object recognition) or Contact Center AI (for obvious purposes).

To make AdaNet easier to implement, the framework relies on a couple of familiar Google APIs, namely TensorFlow Estimator, which helps to store essential information in a single repository, and TensorFlow Board, which basically helps AI teams to create and show an actual live feed of how an ML system is trained.

AdaNet is released under the Apache License 2.0; code samples, scripts and tutorials are available from GitHub. For more detailed insights, please refer to the research paper AdaNet: Adaptive Structural Learning of Artificial Neural Networks, which can be downloaded from the website.


Comments on this article

No comments yet.

Please Login to leave a comment.


Please login

Please log in with your Fujitsu Partner Account.


» Forgot password

Register now

If you do not have a Fujitsu Partner Account, please register for a new account.

» Register now