Some companies are famously, even obsessively secretive about internal development. We never get to see discussion of areas they are working on (other than through patent filings) – we only see the polished and released product/service. Amazon is one such company but Apple must rank for many of us as the pre-eminent company in this class. If you’ve ever had a meeting with an Apple technical team, you’ll understand. They can ask you any technical questions they want but your scope for asking questions is very limited and their answers, if any, will be given only in general terms.
When you’re in the lead or you think you have the special sauce that will push you into the lead, secrecy is an understandable tactic. But when you’re not in the lead, or at least not perceived to be in the lead, a little in-process signaling can help, as in “Hey look, we’re working on this stuff too!” This not only lets the wider world know that you’re not losing your technology edge, but it also helps you recruit. Both of which can be pretty important when other 800lb gorillas have already staked out a domain and you seem to be on the outside looking in. As is the case with Apple and AI, at least as far as the rest of us are concerned. (Sure they have Siri, but that’s old news compared to the continuous PR drumbeat from Google and Facebook.)
Apple announced very recently and somewhat informally that they would change this policy in the AI domain, as least as far as academic publications are concerned. I’m not surprised. If you’re a hot AI researcher with a newly-minted PhD from one of the top schools, where would you rather go – a company with a leading-edge AI program where you can continue to polish your credentials by publishing yet more papers, or a company with undiscoverable AI credentials where you can disappear from view? Kudos to Apple for recognizing that one of the cardinal articles of their faith needed a little loosening up.
The paper itself is interesting and I’m sure a valuable contribution to the field, if perhaps not ground-breaking. The domain is image recognition and the goal is to improve the effectiveness of training using synthetic images (which are already labeled) complemented by similar but unlabeled real data to provide in effect unsupervised training to improve the synthetic images. The intent behind this is to be able to provide much larger sets of labeled training images (since the synthetic images can be generated) without the need for arduous labeling across those sets.