cool hit counter What AI-related technologies are used in Google's first app "Guess the Song"?_Intefrankly

What AI-related technologies are used in Google's first app "Guess the Song"?


In the past two days, Google launched its first AI app "Guess the Song", and the circle of friends has become a stage for soul painters. In a nutshell, you and the AI play a game of "draw and guess" where you draw a doodle on the screen and the AI guesses what you're drawing.

(Image from TechCrunch)

It's actually not the first time Google has launched this product, it's a prototype for a product Google released in November 2016 called Quick, Draw! The web version of the mini-game.

Quick, Draw and this "Guess the Song" are similar to the popular mobile game You Draw Me Guess, but the difference is that Google has arranged for users to play against artificial intelligence instead of humans in these two games.

In Quick, Draw! After drawing a doodle on the board, the AI will always give its guesses, and when six games have been completed, the system will give a general overview, and you can click on each picture to see the AI's recognition results and browse the results of other people's drawings.

The essence of both games is actually an image recognition classification process, where the system cleans and analyzes the features of the graphics to determine what they really mean. Google is releasing Quick, Draw! A video that was shared at the beginning of the game that describes how the game was shaped can help you understand the technical principles behind it in a more visual way: the

Image recognition technology has been an important component of artificial intelligence and one of the research directions that Google has been working on. Good performance image recognition models have important practical functions and can be widely used in many fields such as image retrieval, media content categorization, and video surveillance.

Of course it's not easy to train such an AI, as Google mentioned when introducing "Guess the Song", a neural network driven by Google AI. The network is trained from the world's largest dataset of over 50 million hand-drawn sketches. Back in August last year, Google did a presentation specifically on this dataset, which was also shared by Collective Intelligence at the time

This massive dataset is made up of over 15 million users playing Quick, Draw! The process of interacting with the machine in time is derived and currently contains 50 million doodles. Google has open sourced the dataset, collating the graffiti drawings into an open dataset (which will be expanded to 800 million works later), designed to help developers around the world be able to use it to train neural networks for data analysis, product design, help researchers study the drawing habits of people around the world, and even help artists create novel works -

For those of you who are in need of this, you may want to check out the dataset at

https://github.com/googlecreativelab/quickdraw-dataset

The GitHub repository where the dataset is located also contains some of the developers' own implementations of Quick, Draw! model, which is very informative for those who want to reproduce this technique, don't miss it.

That said, and playing Quick, Draw! In the same way that we play "Guess the Song", we are also helping the AI grow. In Quick, Draw! When it was first released, it used only a few hundred training data and the system's recognition rate was not ideal, but after subsequent engagement and play by up to ten million users and hundreds of millions of doodles to help the AI learn, the system's recognition rate has become quite good and can recognize the same object in different forms (for example, for both the wide-brimmed hat and duck-tongued hat graphics, the system can accurately identify them as "hats").

In fact, this AI is also far from being as simple as recognizing human doodles and playing games with us. For example, through global users on Quick, Draw! On interacting with AI, we will be able to discover the differences in cognitive perspectives and preferences of users in different countries, which in turn will enable us to understand the aesthetic differences in how users depict things around the globe. By doing so AI will be able to make meaningful guidance for localizing our products and apps in different countries and regions, such as how the UI of websites and apps should be designed.

AI can even draw for us after learning people's doodling poses and habits, like Google did with the release of Quick, Draw! Shortly after, a drawing robot called AutoDraw was introduced on top of it that automatically sublimates your doodles into beautiful artistic images. Think about it, designers, you used to need to use the software to design the graphics, but now you can just do the general pattern, AI can automatically help you to transform the desired results, so save time and effort, isn't it beautiful?

(Schematic of the AutoDraw effect)

Plus, by recognizing and learning human drawing and writing habits, the trained AI can help police solve crimes! For example, this year's ICFHR18 conference included a research paper called "A New COLD Feature based Handwriting Analysis for Ethnicity/Nationality Identification," in which researchers analyzed handwriting fonts from citizens of five countries after collecting them, and then used those datasets to train a machine learning model that was able to determine the nationality of the person writing by recognizing the handwriting.

The researchers then selected 100 subjects from Malaysia, Iran, China, India and Bangladesh, had them handwrite the same English, and then used a linear distribution (COLD) recognition tool to analyze the shape distribution characteristics of their written fonts, such as straightness, curvature, etc.

(AI is analyzing the writing habits of people of different nationalities)

The results showed that the subjects' handwriting showed a clear "nationality" even if it was not written in the country's script. Because this is the difference in writing habits created by words, instantly switching to writing another language retains them just as subconsciously. For example, the Chinese, who are used to straight strokes in Chinese characters, will not change their straight writing habits even if they write the Roman alphabet by hand, while the Indian and Bangladeshi, whose scripts are more "rounded", write in a more "round" style. So, if there is a transnational suspect in the case and there is a handwriting left by the suspect in the physical evidence, the AI will be able to analyze the nationality of the suspect.

As the training data becomes more and more extensive, AI will lead to more innovations like this one.

Sure, Quick, Draw! The most direct meaning of "Guess the Song" is to let the public participate in the machine learning process by means of a game, so that AI technology becomes accessible to everyone and everyone can experience the fun of AI.

In fact, Google has always had a great tradition of turning the kinds of projects it has and is working on into interesting little experiments for the average user to experience and test, such as those containing Quick, Draw! The mini-game's project AI Experiments.

(Small experiments in the painting section of AI Experiments)

With these little games and experiments, it's a creative way to give the general public a taste of the latest artificial intelligence technology.

Author: Jingliu Jiji Source: Nuggets

links:https://juejin.im/post/5b559b76e51d45616f4596dd


Recommended>>
1、C Practice C Implementation of Single Linked Tables
2、Docker private repository example
3、Problems related to prime numbers practice c
4、Talking about eurekas delta configuration
5、Open source Hadoop is growing rapidly users and vendors are each getting what they want

    已推荐到看一看 和朋友分享想法
    最多200字,当前共 发送

    已发送

    朋友将在看一看看到

    确定
    分享你的想法...
    取消

    分享想法到看一看

    确定
    最多200字,当前共

    发送中

    网络异常,请稍后重试

    微信扫一扫
    关注该公众号