/* ---- Google Analytics Code Below */

Thursday, December 29, 2016

Apple AI Research for Creating Simulated Training

Interesting premise, to use simulation to produce training sets, a kind of gamification.  We did something similar by simulating forward alternative choices in the supply chain.

Apple delivers on promise by publishing its first artificial intelligence research paper
Apple Inc. has made good on its promise to start publishing research papers on artificial intelligence by publishing its first paper in the field.

Spotted first by Forbes, the paper, titled “Learning from Simulated and Unsupervised Images through Adversarial Training,” was submitted for review in mid-November before being ultimately published by the Cornell University Library on December 22nd.

The research covers the techniques required for training computer vision algorithms to recognize computer generated versus real world images.

According to the research, synthetic image data is often “not realistic enough,” leading the network to learn details only present in synthetic images and fail to generalize well on real images.

The paper proposes that a solution to that problem involves “Simulated+Unsupervised learning,” a process that relies on a new machine learning technique, called Generative Adversarial Networks, that is said to increase the realism of a simulated image by pitting two neural network, generator and discriminator, against each other to discern generated data from real data. .... " 

 ... In delivering the outcome, Apple researchers have made key modifications to the standard GAN algorithm to preserve annotations, avoid artifacts and stabilize training, including a “self-regularization” term, local adversarial loss, and updating the discriminator using a history of refined images, leading to an outcome which is described as enabling the generation of highly realistic images.

In more simple terms the processes described in the research paper help computers see artificially generated images better through improved implementation of artificial intelligence and machine learning.

Apple is expected to publish more research in the year ahead in the fields of “volumetric detection of LiDAR” and “prediction of structured outputs,” both areas the company revealed it was researching earlier this month.

The research paper was written by Apple vision expert Ashish Shrivastava with the support of engineers including Tomas Pfister, Oncel Tuzel, Wenda Wang, Russ Webb and Apple Director of Artificial Intelligence Research Josh Susskind.  ... " 

No comments: