Apple opens digital journal to showcase its machine learning developments

Apple opened a new digital journal to showcase some of the developments it is making in the field of machine learning. In the first entry, it explains what it is doing to help improve the realism of synthetic images, which can, in turn, be used to teach algorithms how to classify images, without needing to painstakingly label them manually.

One of the biggest hurdles in artificial intelligence is teaching it things that humans take for granted. While you could conceivably hand-program an AI to understand everything, that would take a very, very long time and would be nigh on impossible to power. Instead, machine learning lets us teach algorithms much like you would a human, but that requires specialist techniques.

When it comes to teaching how to classify images, synthetic images can be used, but as Apple points out in its first blog post, that can lead to poor generalizations, because of the low quality of a synthetic image. That is why it’s been working on developing better, more detailed images for machines to learn from.

Although this is far from a new technique, it has traditionally been a costly one. Apple developed a much more economical “refiner” which is able to look at unlabeled real images and reference them to refine synthetic images into something much closer to reality.

However, how do you select the correct real image to give the refiner a strong source material to base its refinements on? That requires a secondary image identifier, known as the discriminator. It goes back and forth with the refiner attempting to “trick” the discriminator by gradually building up the synthetic image until it possesses far more of the details of the real images. Once the discriminator can no longer properly categorize them, the simulation halts and moves on to a new image.

This teaches both the discriminator and the refiner while they compete, thereby gradually enhancing the tools as they build up a strong library of detailed synthetic images.

The learning process is a detailed one, with Apple going to great lengths to preserve original aspects of images while avoiding the artifacts that can build up during image processing. It is worth it though, as further testing has shown vastly improved performance for image categorization based on refined synthetic images, especially when they have been refined multiple times.