Apple is a famously secretive company but last year, its head of AI research made a public pledge that on that front, things would be different going forward. Now Apple has made good on its promise by releasing its first publicly accessible paper on AI research.
There was an element of pragmatism to Apple's about-face, as it has struggled to attract top academic AI researchers, who traditionally like to share findings with the greater community. It would be a stretch to believe that Apple has entirely dropped its worries about retaining trade secrets for AI, and a single academic paper isn't enough to fully assess how open it will truly be—but it's a start.
Apple's paper focuses on image recognition, one of the hottest areas in AI given its applicability to technologies such as self-driving cars, security, content management and more.
It's been common practice to train image recognition algorithms using real-world images that have been annotated by humans. Apple's paper proposes a system that uses synthetic images, which can be automatically annotated, while improving the accuracy of the results. Synthetic imagery poses a thorny problem for machine learning systems, as the Apple paper notes:
[L]earning from synthetic images can be problematic due to a gap between synthetic and real image distributions – synthetic 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. One solution to closing this gap is to improve the simulator. However, increasing the realism is often computationally expensive, the renderer design takes a lot of hard work, and even top renderers may still fail to model all the characteristics of real images. This lack of realism may cause models to overfit to ‘unrealistic’ details in the synthetic images.
Apple's system uses a "refiner" neural network to improve the realism of synthetic images, which are then used to train image-recognition neural networks, with "state-of-the-art" results, the paper states.
Overall, Apple's paper underscores a key requirement for AI moving forward: The availability of large, well-vetted data sets, as Constellation Research VP and principal analyst Alan Lepofsky writes in a new report:
AI-enhanced software requires massive training sets to learn from. If doctors are going to use AI to help with a diagnosis, they want the application to learn from thousands of medical journals and patient histories. If financial brokers are going to use AI to help make trades, they want the application to know everything about the last several decades of the stock market. If employees want AI to help their productivity, they need the application to learn from thousands of data sources, including emails, calendar entries, documents, and social media posts.
Of course, the mere presence of large data sets isn't enough; the mapping of connections within data is hugely important for AI in providing more accurate insights and predictions, Lepofsky adds. Perhaps 2017 will see Apple's AI researchers release work focusing on the latter area.
24/7 Access to Constellation Insights
Subscribe today for unrestricted access to expert analyst views on breaking news.
