Apple first published research on artificial intelligence

Earlier this month, Apple announced that it would begin publicizing its research on machine learning and artificial intelligence (AI). This will help uncover the secret of one of Apple's most secretive projects.

Now, after a few weeks of waiting, Apple finally released its first AI document. According to Macrumors, the document released by Apple is titled "Learning from simulated images and without control through reverse training". The main content of this document is focused on Apple's efforts in the field of image recognition min .

Specifically, in this document, Apple describes a program that can decode and understand digital images like Siri Intelligence and identify faces in Photos apps of iOS 10, but in one higher level.

Picture 1 of Apple first published research on artificial intelligence
The main content of this document is focused on Apple's efforts in the field of intelligent image recognition.

In the first part of the document, Apple paid much attention to the advantages and disadvantages of using actual images to compare with using artificial images in training AI. Using realistic images will require a "costly and time-consuming process" as well as a large amount of manpower to label (annotate) each individual photo. Meanwhile, if using CGI technology to create and customize artificial images, the labeling process will be much simpler because "annotations can be automatically added".

This has led to widespread use of artificial images to train AI in recent times. The use of artificial images saves costs very much and is easy to customize and available.

However, the abuse of artificial images will lead to a deterioration of the image recognition program. The reason is that "artificial images are often not practical enough". This leads to users receiving only results on an artificial image, instead of the generalization of an object in reality they expect.

To solve this problem, Apple has proposed a combination of artificial images and realistic images in the "opposite training" process to create the most advanced image recognition AI program.

Picture 2 of Apple first published research on artificial intelligence
Diagram incorporating artificial images and unbranded photos of Apple.

"In this document, we propose simulated and uncontrolled learning to enhance the authenticity of artificial images created by simulating actual unlabeled images. Improvements The authenticity allows us to train machine learning models on a larger database without the need for human help in data collection and labeling, " Apple writes in material.

At the end of the document, Apple delves into the rest of the image recognition program like the process and how AI backs up new findings. Apple also hopes to train AI soon to be able to identify videos.

At the AI ​​conference held in Barcelona (Spain) a few weeks ago, the head of Apple's machine learning department was Russ Salakhutdinov and some colleagues discussed various topics such as the signs of health, predict the structure of outputs, smart assistants and activity identification. After image recognition, Apple will continue to publish its AI studies on other topics in subsequent documents.