Česká verze příspěvku

While reading news stories on research or products involving deep learning, I get often surprised how inaccurate and misleading the news stories are. It is probably a problem of almost all expert fields which happen to appear in media, luckily they do not bother me as much as AI.

News stories compete the get our attention in a world with so many things to do. They need to win our attention over our work, Facebook posts, friends, family, books you want to read—and that is not an easy task at all. Media scientists claim the news must satisfy some criteria to even get a chance to attract the readers’ attention, these are called the news values. In case of AI news, these these are that technologies have a stable place in our culture (cultural proximity), everyone knows what technological progress is and what it is good for (unambiguity) and possibility to personalize the news stories, both through stories of its developers and stories of the technology users.

The news must be simplified and put in a shape that fits the news values. Media indeed need to avoid complicated technical terminology and replace it with something that is more familiar to the readers—in this case, the terminology and metaphors often come from science fiction literature.

Unspoken connections between emerging technologies and sci-fi genre of course make the news stories more attractive and help to make the technologies more popular among the public. On the other hand, it raises false expectations both in terms of possible use and misuse of the technologies. Public discussion is misdirected towards hypothetical problems and overlooks the real ones.

Google’s Machine Translation Developed its own Language

At the end of 2016, Google published a study showing how their neural machine translation models can be modified in such a way that a single model is capable of translation between multiple language pairs at the same time. Moreover, it is supposed to work not only for the language pairs it was trained for, but also for language pairs that were never seen during the training, as illustrated on a scheme from Google Research Blog.

Google's Zero-Shot Translation

Nowadays, we need a separate model for every single language pair, and for some of them there is only little bilingual texts that can be used for the model training. Therefore, if the proposed model worked properly, it would be a major advance in machine translation. However, this did not happen. The results are interesting from the theoretical perspective, but the performance of presented models is by large margin worse than what are users of Google Translate used to.

Many news servers called the input representation the models had to learn as interlingua. It is a hypothetical meaning representation that should be the same for all the languages, an ultimate analysis of a sentence (which I doubt is even theoretically possible). Some news even claimed it created a new language. This news has been reported even by the most read technology news servers:

The articles not only do not mention that the translation was useless for any practical purposes, but they also make a false impression that the intermediate representation that the system uses is something that can be used a language. News stories like this tell the readers that current neural networks are so intelligent that they can not only use human language, but they are capable of inventing a new, presumably better language. Is is an ability that we would normally attribute only to an autistic genius and a machine with such capabilities sounds like a scary and unpredictable thing. Nevertheless, the representations the system uses are tables with thousands of real numbers for which we have no direct interpretation. It has none of the properties people usually attribute to languages.

Facebook’s AI Grew out of Control, so they Had to Stop it

Most of the current models for natural language processing model only how the language looks like under different circumstances instead of how to use it. For instance, machine translation models do not bother with that all. So far it seems, it might be enough to teach the model how does a sentence in a target usually look like for the given source sentence, in order to translate it (almost) correctly.

If we wanted to create a program that negotiates with someone about something, a program that intentionally follows the goal you want it to achieve, trying to simulate what people do in similar situation does not seem like a good strategy. Obviously, you need to know what you want to achieve in order to achieve it. You cannot do it just by mimicking of what people usually do when they negotiate.

During the last summer, a research team at Facebook did an experiment whose goal was exactly this. In the experiment, they trained chatbots which were supposed to negotiate with each other about exchanging hats, balls and books. They used the same principle as was used for instance while training the AlphaGo, the first system that beat humans in the game of Go. AlphaGo was trained by playing millions of games against different version of itself and improved through trials and errors in those games. In case of Facebook’s experiments, the chatbots were given some prior knowledge how users chat with each other. Their starting point was in fact knowing how the language look like. Then, the chatbots were improving their negotiation skills by constantly communicating with each other. As a result, the they learned to negotiate efficiently with each other while totally diverging from what they knew at the beginning. The code they have developed for that did not resemble English at all.

The experiment got surprisingly high media coverage. Most of the news stories were telling that Facebook had conducted an experiment with artificial intelligence that had got out of control and therefore they had to stop it.

This time, it were the technology news servers that tried to calm down and explain the exaggerated news from other media:

If we think of it more deeply, it is actually no surprise that the code which the systems had developed was totally different from any human language. After all, human language is probably not the most efficient code for negotiating about hats, balls and books. It has many other fascinating properties: we can write poems or tell jokes using it. We can also see the result of the experiment as an argument, that talking business is not the constituting function of language, which is at least for me good news. Facebook stopped the experiment not because it panicked how dangerous it was, but because it had already delivered results they wanted and it was useless for any practical use.

What are the Takeaways?

The way media talk about AI technologies (including using term AI) often resembles the way in which AI is depicted in the sci-fi literature and movies where the term means something entirely different. The problem is that while talking about the potential problems of AI, the sci-fi conceptualization tends to make us think about sci-fi threads. As we saw in these examples, media tend to do it whenever there is a story that fits into this framework, so they can come up with an attractive news story. What comes to our minds are technology getting out of control eventually exterminating humanity or a supervillain using the emerging technology to conquer the world.

Don’t worry, none of it is in preparation, neither in Facebook nor Google labs, neither anywhere else. After all, how could a model that estimates a conditional probability of words in an English sentence given a sentence in a different language conquer the world? The technologies are indeed going to have a big effect on society and this should be publicly discussed andp not shouted down by sensational speculations.