Apple AI has officially released its first research paper after the company announced, earlier this month, that it will be sharing its results.
The world of AI is growing at a faster than ever rate. AI or Artificial Intelligence is considered by some to be amongst the future of technology. Many big industry names are involved in its development.
An alliance has already been formed. It involves companies and projects such as Amazon, Facebook, or the Google DeepMind. However, there was one notable absentee.
Apple is also one of the big companies to be developing AI technology. Still, until just recently, the company was quite closed off. It did not allow its researchers to publish research papers on the matter.
However, in early December, the new head of Apple AI issued an announcement. Russ Salakhutd revealed that the company will now be more open in its research.
And he has gone and kept his promise. The first such paper was released earlier this week.
The study was submitted on December 22. It has been made public on the arXiv preprint server. The Apple AI research paper is titled as follows.
“Learning from Simulated and Unsupervised Images through Adversarial Training”. It is centered on intelligent image recognition tools and technology.
More exactly, it outlines the abilities of such a technology. A program using it could come to recognize computer-generated images. Not only that, but it should also decipher them.
Apple AI researchers released some details on the matter. According to them, such a technology has yet to be achieved. The reasons were also outlined.
As such, current synthetic data is usually not realistic enough so as to be read. The easiest solution, increasing realism, is in its part computationally expensive.
Graphics has recently registered progress. According to the researchers, it has become more tractable to use synthetic images in training models. By doing so, studies could potentially avoid the expensive annotations.
However, there is still a gap between such images and real image distributors. As such, synthetic image learning may not lead to the desired performances.
In the current paper, the researchers propose a solution. They are drawing attention to a S+U learning. This is a Simulated + Unsupervised method.
Its goal would be to improve synthetic image realism. It would do so based on unlabeled real data from a simulator.
Apple AI researchers developed the SimGAN. This is based on S+U learning which also uses adversarial learning. The latter uses two competing neural networks. These basically compete against each other.
In this case, one is a real image generator. The other is the discriminator, which tries to distinguish in between the real and the generated images.
The such developed neural network was termed a “refiner network”. This is continuously trained with an adversarial loss.
As such, they are trying to fulfill the S+U learning algorithm requirement. This is adding realism.
The current is the first of the quite many more expected ones. As the Apple AI has opened its gates, specialists are expecting other researchers published and authored by them.
Industry reports issued a few theories as to the new openness. Some suggest that the company is trying to make up for its lost ground. Other market competitors have marked significant advances in the AI area.
By creating a more open environment, Apple could also be trying to bring in new talents. This would ensure a further development as it would bring in fresh ideas.
Image Source: Pixabay