The heart of the machine is original.
Everyone’s favorite Tumblr has gone down…
In early December, Tumblr announced that it would ban all adult content. The new rules will take effect on December 17. This behavior seems to be taking the initiative to say goodbye to “30% of Internet traffic”, long-time Tumblr drivers sent a song “cool cool” to it.
Tumblr has been known (and possibly notorious) for allowing NSFW content since it launched in 2007. Is this a passing fad by the company’s management or is it a case of a “clean SLATE”? Regardless of Tumblr’s intentions, adding moderation mechanisms to social networks to filter pornographic images/videos requires an investment of technology and manpower.
In Tumblr’s official announcement, CEO Jeff D ‘Onofrio said, “… [Tumblr] will ban adult content using industry-standard machine monitoring and increased human monitoring.”
But computers that just ‘open their eyes’ aren’t perfect. “Wired wrote in a recent article. Many Tumblr users have complained on Twitter that its authentication system is plagued by misjudgments. According to Wired, Sarah Burstein, a professor at the University of Oklahoma Law School, was flagged for just Posting a few patent drawings for her design. The article says this is not only inefficient, but also hurts users’ feelings. Many Tumblr users flocked to Twitter in frustration, with the New York Times describing them as “jumping ship.”
Tumblr has had trouble accurately identifying NSFW content for years. Yahoo bought Tumblr for $1.1bn in 2013 and Verizon acquired it four years later. Tumblr was owned by Verizon subsidiary Oath and soon after launched “safe mode”, which automatically filters adult content into search results.
Although artificial intelligence can process a large number of pictures at the same time, but after all, it is not human, there will inevitably be some funny mistakes. Microblogging platforms like Tumblr, in particular, have a complex user base, and the line between porn and non-porn can be tricky to navigate. “If the standards are too strict, content such as photography and art may be identified as pornographic and harm the user experience, while if the standards are too low, pornography will continue to spread,” said Jiang Zelong, director of product operation at Tupu Technology. In addition, there are various types of content on the platform, including text, pictures, video, live broadcast, etc., all of which have different real-time performance and greatly different auditing standards, which undoubtedly put high requirements on development and operation teams.”
Tupu Technology is an early AI startup in China that uses deep learning technology to provide image and video content review services. On issues such as the censorship of pornography on the platform, Tupu explains something to the Heart of the Machine.
What is the process of image yellow system?
Yellow discrimination system workflow is this: the first is to model, followed by the classification standards for pornography, then collected a large amount of material, is classified and labeled with these marked with good material for training, finally let the machine to study the inside of the various classification characteristics, constantly adjust the parameters of the model and eventually get the best recognition model.
When the machine recognizes the image, the image data will be converted into digital information and brought into the model for calculation. According to the calculated value, the image will be marked as “normal, sexy, pornographic” three categories.
Such a system would also involve humans because of the imperfection of “recognition”. After identification, the system will tell the user the judgment result and probability, and the user will do corresponding processing according to the result, such as automatic deletion or manual access review. If machine recognition error is found in manual review, data learning will be carried out on the pictures of the same scene and parameters will be adjusted until the error rate reaches the lowest value.
The technical core of AI huang jian is Deep Learning theory. In layman’s terms, deep learning can be understood as a blank brain, and massive data is the experience injected. When we put a lot of porn, sex appeal, the normal sample attribute tells the depth study of the engine, leave the engine to keep on learning, then they do the right reward, do wrong to punish, of course, these rewards and punishment are mathematically, finally will learn a blank head connection model, this model is to identify sex and not porn.
Deep learning is an Artificial Neural Network (ANN). To understand ANN, let’s take a look at how the human brain works.
The figure above shows the process of understanding visual information from the outside world. From the Retina, edge features are extracted through the low-level V1 region, to the basic shape or local target of V2 region, and then to the whole target of the high-level (such as judging a face), and to the PFC (prefrontal cortex) of the higher level for classification and judgment. In other words, high-level features are the combination of low-level features, and the expression of features from low-level to high-level is more and more abstract and conceptual, that is, more and more capable of expressing semantics or intentions.
Deep learning is precisely to form more abstract high-level features (or attribute categories) by combining low-level features, and then obtain a high-level expression based on these low-level expressions through linear or nonlinear combination. And it’s not just images that have this pattern, it’s also sound.
Now let’s look at a simple model of deep learning.
One of the main advantages of deep learning is that it can make use of massive training data (i.e., big data) to continuously improve the identification accuracy in the learning process, but it still requires a high amount of computation. In recent years, thanks to the improvement of computer speed, the rise of large-scale clustering technology, the application of GPU and the emergence of numerous optimization algorithms, the training process that takes months can be shortened by days or even hours, and then deep learning can gradually be applied to industrialization.
For the development team, the difficulty of making products in this field lies in how to obtain large-scale annotated data, integrate the computing cluster with GPU, and adjust parameters for their own projects. The team needs to continuously input new data and iterate to improve machine recognition accuracy.
What if you have pictures and video?
In a world where short videos are popular, tech companies are also dealing with a huge amount of video content. Pictures are static, video/live is dynamic, and the complete review of video content includes the review of pictures, text and voice, so it will be more complicated. Take the video picture audit as an example. When identifying video and live broadcast, dynamic content can be decoded into picture frames first, which is similar to the method of identifying static pictures.
In addition, the scene and characters in the live broadcast vary greatly, and the audit requirements are strict, so the identification difficulty is relatively high. It is necessary to carry out real-time frame capture transmission identification of the room, and realize the early warning processing by combining with manual. The overall picture quality of videos is worse than that of pictures and live broadcasts, which will affect the recognition effect to some extent. Usually, screenshots are taken at equal intervals in the unit of videos, and the results of multiple screenshots of a video are used to comprehensively judge whether a video is pornographic or not.
If enterprises identify every frame of video or live broadcast, the amount of data will become huge and the operating cost will be high. Faced with this kind of situation, it is generally used to process the video frame extraction. For example, for a one-minute video, 6-15 frames of pictures can be extracted according to the time period for recognition processing, so as to reduce computing costs.
Can audit be completely machine dependent?
In view of the problem of “manslaughter”, which is often mocked by people, Tupu believes that the accuracy of AI algorithms should be improved while still relying on human beings to make the final judgment. There are two main types of misjudgment: misjudgment of pornography as normal content, and misjudgment of normal content as pornography.
1) Pornography judgment is normal: in dim light scenes, or scenes with large background interference at a long distance, as well as in the case of special interference, it is possible to cause misjudgment; Normally dressed but in fact dewy, obscure movements and postures.
2) Normal judgment of pornography: naked clothes but no dew point in reality, objects resembling sexual organs, close to pornographic movements but not in reality (such as holding stick objects, hands on sensitive parts normally), etc.
Machines can help enterprises greatly improve audit efficiency and accuracy. Taking the yellow authentication system of Tupu Technology as an example, it can review nearly 1 billion pictures every day, with recognition accuracy higher than 99.5%, and can save more than 95% of audit manpower for enterprises. But at present, even in a long time, artificial intelligence can not completely replace artificial yellow detection. This is because machines still have a hard time understanding the meaning behind the content and do not switch between different cultural scenes. Therefore, machine + manual audit method is recommended.
Simple algorithms and models can train the machine to make a completely correct judgment of the situation, but in practical application, the machine has no independent thinking and its own subjective consciousness, so it still needs manual assistance for confirmation. For example, the picture provided by the customer is too fuzzy or the light is too dark, and the training data cannot be fully covered, and other objective reasons, the machine can not produce a high score confirmation picture, which requires manual assistance.
It seems that the AI image recognition system can use the existing, “yellow division” is not please. After announcing a ban on adult content, Tumblr’s app has finally reappeared in Apple’s App Store. Where does the car end up going? Let’s wait and see.