Google's "AI Guess the Picture" app is popular on WeChat Moments: user participation makes it smarter

One day after its launch, Google's first WeChat mini-program Guess the Painting app has swept some people's Moments. A large number of "soul painters" are happy to share their paintings in Moments, while also raising many questions: "Why can't AI guess my paintings so well?"

Judging from the communication effect, the first WeChat mini program launched by Google can be said to have achieved great success.

Guess the Picture Song is a very simple app. The specific gameplay is that users team up with Google AI to sketch within a limited time. In each round of experience, users need to sketch a picture of daily necessities (such as a dog, clock or shoe) within the specified time, and AI teammates need to guess the object in the picture before the time is up.

When the AI ​​guesses the object in the picture, the user will enter the next round, and the user's guessing streak will increase accordingly. Users can also invite friends or family to challenge the longest winning streak, share interesting sketches, and continue to collect new questions and pictures during the experience.

The ranking list provided by Guess the Painting Xiaoge shows that a user at the apprentice level of a painting studio is currently ranked first in the ranking list because he drew 288 paintings in a row that could be accurately recognized by AI within 25 minutes and 30 seconds.

The results of the Guess the Picture Song may not be directly proportional to your drawing skills. In fact, many users will encounter such a situation during the experience: they have clearly drawn a certain object very vividly, but the AI ​​just can't guess it. To answer this question, you must understand the operating principle of the Guess the Picture Song.

Guess the Painting Xiaoge uses neural network technology to identify user paintings, which is based on learning from a large number of graffiti samples.

Google gave a brief introduction to it in the official description: As a type of AI technology, computer vision technology allows computers to directly understand the input visual information and thus "see" the world. This technology can not only identify your friends in video calls so that they can add corresponding tags, but also help identify early symptoms of diabetes in human fundus diagnostic images. Thanks to the "neural network" technology, some things that seem extremely difficult can be handled very well by computers: for example, rough sketches can be used to identify what an object is. Now, computers can "see" that a random doodle of ink with ears may be a panda.

The specific support for this game is a recursive neural network called RNN, which can use its internal memory to process input sequences of arbitrary time order, making it easier to handle non-segmented handwriting recognition, speech recognition, etc. The so-called input sequence refers to the stroke sequence when writing, and the before and after process when drawing.

In fact, as early as 2016, Google launched a web version of Quick Draw with a similar gameplay: it will randomly display a noun and ask you to draw it within 20 seconds. Players need to simply outline the object with the mouse, and then Quick, Draw! will judge whether your drawing is similar to the object.

Although the game looks simple and easy, it is an important part of a series of AI experimental tools by Google, which actually uses a neural network algorithm to judge the player's graffiti. Google is trying to use it to study how to let AI learn image recognition and optical character recognition - both of which are core topics in the field of AI.

A year later, Google released the dataset behind the project. The Quick Draw dataset is a collection of 50 million drawings, divided into 345 categories, from players of the game Quick, Draw! These drawings are timestamped vector images and annotated with some metadata, including what the player was asked to draw and the country the player is from.

The guessing picture song that has been flooding the WeChat Moments this time also has a similar purpose: to give everyone the opportunity to experience human-computer interaction driven by artificial intelligence technology.

In the process of continuous interaction with AI, players can see the process of AI guessing pictures. And Google has also thoughtfully made such a setting. When users pass a level for an item, they can see how others depict the item. When a picture data set of an item is displayed in front of you, you may try to understand what "ice cream" looks like in the eyes of AI.

More importantly, every user who has experienced the game of Guess the Painting is actually constantly training the AI ​​through their own paintings based on their understanding of the world.

As Google said when releasing the quick draw dataset: This is not only a huge amount of data, but also an interesting mirror - it can be used to see how to allow the public to participate in training machine learning systems in the form of crowdsourcing, and how to create a dataset that reflects different cultural backgrounds and concepts.

This sentence can also be used to describe Guess the Painting Song: it is not only a small program, but also a way for ordinary people to participate in training machine learning systems.

Judging from Google's move to develop mini-programs, this international technology giant is working hard to explore and find ways to adapt to the habits of Chinese users to participate in the research and development of its own products and technologies.