- Labeled data is very expensive. Historically attempts to learn on synthetic data has failed because ConvNets are very good at detecting small visual artifacts in the synthetic data and using those for classification during training. At test time on real data, those artifacts aren't present so model fails. A technique that can beat state-of-the-art (admittedly on a very narrow Eye Gaze dataset, but still) by only training on labels from synthetic data and testing on real data is important.
- They present a useful new idea to improve GAN training: using a history of "fake" images, rather than only the latest fake images from the generator. Ask anyone who has tried to train a GAN: the training is really unstable, each network only cares about beating the latest version of its "opponent". They show good improvements by saving many previous fake outputs to make the generator more robust. This reminds me of Experience Replay from DeepMind for RL.
- It's a published paper from Apple! Great that they are starting to contribute back to the research community.
As context, this is Apple's first published AI research paper. After Russ Salakhutdinov was hired, he promised that Apple would start publishing AI papers in peer-reviewed conferences... and this is them making good on that promise.
I work in AI (using HealthKit data, not as part of Apple), and my quick read is that although this particular paper isn't especially groundbreaking, it's a good omen for things to come. The idea they propose is a good twist on Generative Adversarial Networks—which have been around either since 2014 or the 1990's depending on who you ask—but you might see dozens of similar papers at a conference like NIPS.
That said, Apple is using state-of-the-art techniques, submitting to peer-reviewed conferences, publishing results on well-known baselines, AND publishing openly, none of which were true of them even a couple of years ago. This is real research.
So kudos to them. I'm really curious to see how they publish in core product areas like Health or Siri -- they could make some major contributions to research in addition to products.
I found the paper very well written. It's a testament to Andrew Ng's ML class that you're able to follow a state of the art paper (terms, equations, conventions) after his lessons.
The notion of a "visual Turing test" was interesting (can humans distinguish real from simulated images). The sample sizes (and p-values, for all their issues) were fairly small (10 participants, 100 images each).
I wonder if there are volunteer participant efforts (similar to HealthKit) to help automate these types of evaluations. You could even have a Recaptcha-style service to show your simulations are realistic enough.
Yay! Apple publishing ML papers. The paper's pretty interesting in concept, but I would have loved to see more variety of samples.
If we could make rendered images more natural, wouldn't that have a huge benefit on gaming? We could make games almost identical to real video/images. Couple that to 3D VR headset and you have a pretty sweet system.
- Labeled data is very expensive. Historically attempts to learn on synthetic data has failed because ConvNets are very good at detecting small visual artifacts in the synthetic data and using those for classification during training. At test time on real data, those artifacts aren't present so model fails. A technique that can beat state-of-the-art (admittedly on a very narrow Eye Gaze dataset, but still) by only training on labels from synthetic data and testing on real data is important.
- They present a useful new idea to improve GAN training: using a history of "fake" images, rather than only the latest fake images from the generator. Ask anyone who has tried to train a GAN: the training is really unstable, each network only cares about beating the latest version of its "opponent". They show good improvements by saving many previous fake outputs to make the generator more robust. This reminds me of Experience Replay from DeepMind for RL.
- It's a published paper from Apple! Great that they are starting to contribute back to the research community.