The picture is quiet and AI is more than it, MIT Black Technology makes the picture seconds smaller

Have you ever thought about it? When you show you any photo, you may see more than just a still image, but a smart "little video." Today, with the help of machine learning, the next series of actions can be predicted from still photos, and the accuracy is still quite high.

Whether it is beauty riding, dog taking Frisbee, or someone suddenly falling, etc. Imagine these continuous actions is one of our most basic skills, we do not need to consider a lot of information for prediction, such as gravity, inertia and falling Instinctive reaction. Then, the ability to let computers learn such predictability is undoubtedly a key challenge in machine vision.

Researchers from the Massachusetts Institute of Technology are working hard to solve this problem. They have shown a series of very impressive results. Using specially trained neural networks, images are converted to video and the computer predicts what will happen next. However, their models still have many limitations. Video is usually only a few seconds long, the files are small, and the images are often confusing. But this is still an impressive innovation in machine imagination. Computers have taken a step forward on the road to understanding the world like a human being.

Train this neural network using more than 2 million video clips downloaded from Flickr. All scenes are divided into four types: golf course, beach, train station and hospital. This set of footage is very stable and eliminates camera shake. With this data, the team's neural network can not only produce short videos that resemble these scenes, but also produce continuous pictures based on a still image. This essentially prejudges the actions that will happen next, but the effect is very limited at present. It can only infer the changes in the pixels, rather than based on the understanding of the entire scene.

Here is the rendering:

Here, we can see the effect after implementation. For example, on the beach, you can see the fluctuations of the waves; at the train station, the prediction model predicts that the train will run. However, when asked to predict how someone is going through the golf course, the results seem distorted and the images are blurred.

The researchers mentioned that computer predictions are often not in line with normal logic, but at least their judgment of the trajectory is reasonable.

Machine learning systems have made many advances in related fields, including predicting handshakes and hugs, and even generating audio that matches video. Yann LeCun, head of Facebook's AI department, mentioned this topic in an interview last year, saying that the pre-judgment trajectory is an important part of the development of predictive computers. However, to really understand the video or image, and the actions that may occur next, it takes more research staff to spend.

"If you were looking at Hitchcock's film, then I asked, 'What will happen to the movie's plot after 15 minutes from now?' You must try to predict who the murderer is."

LeCun said: "To completely solve this problem, you need to understand this world and human nature. This is where the real fun lies."

The ability of artificial intelligence in forecasting has become stronger and stronger. However, in order to achieve more accurate, natural, and practical results, more sophisticated models are needed. Researchers may need to consider more factors, build more complex neural networks, and use more data sets to train models. Only in this way can it be possible to truly predict the continuous motion in the image through machine learning techniques.

Via the verge