Facebook, Inc (NASDAQ:FB) has made its latest statement outlining that the billions of images on Instagram will help it in the building of the sophisticated AI deep learning models. The social media giant has provided a detailed account of the large number of photos that were annotated by users with hashtags. It added that it had moved ahead to use the data in the training of the image recognition models.
The GPUs running around the clock were found to be quite useful especially when it got to parsing data. The interesting bit was that Facebook managed to attain 85.4 percent accuracy on ImageNet thus beating the industry benchmarks.
The pre-training research
A person well conversant with the latest developments but who wanted his identity kept anonymous said that ‘pre-training’ research was more about coming up with systems that would help establish relevant hash tags. It was about identifying the particular hashtags which were fundamentally synonymous. The other thing was about prioritizing the more specific hashtags and of course the general ones could wait.
The ‘large-scale hashtag prediction model’ is the term that the research group used in reference to the latest development. Several analysts have come up to term the associated privacy implications rather interesting.
Mike Schroepfer thoughts on the latest developments
THE HERALD FINANCE REPORT
Start your workday the right way with the news that matters most.
Facebook’s chief technology officer Mike Schroepfer opined, “We rely almost entirely on hand-curated, human-labeled data sets. If a person hasn’t spend the time to label something specific in an image, even the most advanced computer vision systems won’t be able to identity it.”
In its latest statement, Facebook outlined that it recognized the useful applications as well. It exuded confidence that the use of hash tags as labels for computer vision would do much in terms of impacting the way it ranked images in feeds as well as towards the enhancement of the AI system’s understanding of video footage.
With the passage of time the training data sets are growing bigger and most probably in the near future there will be the need for unsupervised learning.