In order to create a convenient and powerful AI machine learning model, Facebook uses 3.5 billion public photos and countless hashtags on Instagram as the material for training the model.
Training machine learning to recognize objects in pictures often requires manpower and time to mark names and categories, so that the machine can be trained smoothly. In order to create a convenient and powerful AI machine learning model, Facebook has a super secret weapon ─ 3.5 billion pieces of Instagram exposes photos, and countless hashtags.
Instagram public photos, into free training model material
If you’re in the habit of adding hashtags to photos when you post, chances are you’re helping Facebook train a machine learning model.
In the past, it took a lot of manpower to train a machine to learn to recognize objects. For example, if you want a computer to understand the "mug" thing, the human must first mark the thing as a "mug", and then the machine can have a basis. Learning, and now Facebook has found a way to train machine learning with a small amount of human effort.
At F8, the developer conference the next day, Chief Technology Officer Mike Schroepfer said that Facebook researchers and engineers, using the platform's 3.5 billion public photos and more than 17,000 hashtags as the basis, Hundreds of GPU parsing data are used to train our own image recognition model, which can now achieve 85.4% accuracy on ImageNet (image database).
But there are also many challenges in the process. Mike Schroepfer said that many users often use the wrong label. For example, although users mark the photo with "flower branch", it is actually "squid". In addition, Facebook must also train those Hashtags that are synonyms, and sort the specifically described Hashtags better than other Hashtags of the same type, and eventually become a set of "large-scale Hashtag prediction modules" to train image recognition modules.
Facebook uses Instagram’s 3.5 billion public photos, plus user-annotated hashtags, to train machine-learning models.
Hashtag training model, will there be privacy issues?
However, at this stage, everyone is more concerned about the "privacy" issue. When a user posts an Instagram photo, can they clearly know that they are providing information to train a deep learning model?
Facebook said that the image data will only be used to identify the association with the hashtag. The data used is public, and even if the photo itself is not tagged, no matter how powerful the machine is, it cannot identify the object in the photo. Simply put, uploading public photos on Instagram, whether users want to or not, is contributing to Facebook's deep learning technology.
Chief technology officer Mike Schroepfer (Mike Schroepfer) said that the process also faced a number of challenges, such as many users often use the wrong label.
But hashtags are still the key to training. If users don’t want their photos to become material for training machine learning, don’t add hashtags to public posts. Facebook emphasized that the team will only capture material related to the image in the public hashtag, and will not infer user behavior from the content of the photo.
Facebook's method of constructing and training machine models is more interesting than identifying the accuracy itself. It organizes huge and messy materials into neat and well-founded materials. In the future, it can help users search for information more accurately and quickly, and even use AI to automatically generate pictures. Explain and assist in reviewing the content of the platform.
Article source: https://www.bnext.com.tw/article/48994/facebook-trained-image-recognition-ai-instagram-pics