(RNN) – Using hashtags on Instagram can help spread your picture far and wide, and attract the attention of other accounts who share your interests.
It also might surprise you to learn it’s been helping Facebook develop artificial intelligence technology.
Facebook revealed some of their findings at their annual F8 developer conference on Wednesday from a research initiative focused on advancing so-called deep learning methods in image-recognition technology.
Deep learning loosely describes the process by which extremely powerful machines and networks are effectively designed to function more like a human brain. Rather than the pre-programmed processes a typical computer performs, the idea is that deep learning systems analyze information and apply reactive processes.
This, however, requires massive amounts of data to feed to the systems to analyze and learn from.
To achieve what the company called a “genuine breakthrough” in image recognition – which, for instance, is crucial to how self-driving cars are being developed – Facebook drew from billions of publicly available Instagram posts. And hashtags were the key to it all.
Most data sets fed into machine learning networks require a remarkable amount of human legwork.
For example, in order for a self-driving car to recognize a stop sign, its operating system first needs to learn what that is. That requires feeding many, many examples of a stop sign into the system, and before that can happen a human has to identify and tag images of stop signs to do the feeding with.
This is referred to as “supervised training.”
Facebook’s research team broke though that ceiling with “weakly supervised training,” harvesting the countless publicly available images on Instagram that have already been tagged.
Advancing state-of-the-art image recognition with deep learning on hashtags: new SOTA on ImageNet (1-crop, top-1 85.4%, top-5 97.6%) by pre-training on 3.5 billion images and 17,000 hashtags, by Facebook AIhttps://t.co/Yl8epaXube pic.twitter.com/tyNtSwRGmI
— Alexandr Kalinin (@alxndrkalinin) May 2, 2018
While the messiness of Instagram tags created a far rawer data set than those prepared by researchers, Facebook’s team reported in a paper that they found “without manual dataset curation or sophisticated data cleaning, models trained on billions of Instagram images using thousands of distinct hashtags as labels exhibit excellent transfer learning performance.”
The company said it used a data set with 3.5 billion images and 17,000 different tags in this research.
It reported its computer vision system scored an 85.4 percent accuracy on a measuring tool considered a popular standard, ImageNet. It is a record score.
“These are foundational improvements to image recognition and object detection, representing a step forward for computer vision,” the company said in a blog post.
Copyright 2018 Raycom News Network. All rights reserved.