How Facebook Taught Its AI to Speak More Than Fifty Languages

With the advent of movies like The Terminator and War Games the idea of machines turning against their creators embedded itself in society. However, at that time there was no technology that could enable computers to learn quickly and effectively. Nowadays, artificial neural networks present humanity with new possibilities... and new dangers.

In the 1990s, with the appearance of search engines such as Yahoo and Altavista, many users required machine translation, for which they used services like Babelfish. Computers at that time mostly relied on dictionaries which used simple "if-then" rules and high-level object-oriented logic, which made it difficult for them to learn new things.

But a decade later, new social media platforms like Facebook have gained massive following among Internet users, who started sharing increasing amounts of data over the network, and that often required new methods of machine translation.

Social media opened a new door of opportunity for computer scientists, who were working on sub-symbolic machine learning algorithms.

In 2011 Facebook staff started feeding the company's AI with user-generated texts and images to teach its neural network new skills. Facebook representative Alan Cooper told Sputnik that currently the social media platform is capable of working with more than 50 different languages.

We actually run tons of AI and machine learning algorithms every day on Facebook. Some of those, they are most used every day actually in translation. We have over 800 million people using translation once every month. And then we also have over 2 billion translations happening every day on facebook.com.

Artificial neural networks are a method of computing where neural units, or artificial neurons loosely imitate the human brain to solve problems. Modern neural network projects usually work with a few thousand to a few million neural units and millions of connections. The method has been used in areas like computer vision and speech recognition, where it is hard to apply regular rule-based programming.

Facebook's handling of user content sometimes has led to criticism from online privacy activists.

In October of 2016 Facebook Messenger and WhatsApp, which is also owned by Facebook, were named the most secure chat platforms by Amnesty International. Other human rights watchdog – Electronic Frontier Foundation — didn't agree and said that there is a number of privacy issues in both programs. However, the company's representative Alan Cooper says that Facebook does a lot to educate users about privacy decisions on the platform.

We heavily invest and believe in privacy. We have an entire privacy center on Facebook that makes sure that every user is very able and educated on what privacy decision they are making on a platform. And in terms of stuff we do we have more of a signal test instead of looking at every single word and understanding every single thing.

Currently Facebook's computers handle more than 1800 different language permutations – a number which is very hard to achieve even for huge and well-organized groups of human translators. And even though the quality of machine translation is sometimes very low, with the input from Facebook users and with additional tuning of the code, it appears that computers are quickly learning from their mistakes. So when it comes to translation, artificial neural networks may eventually replace humans, it’s just unclear when it's going to happen – a year, five years or a decade from now.

We'd love to get your feedback at [e-mail].

Have you heard the news? Sign up to our Telegram channel and we'll keep you up to speed!