Neural networks learn from people to paint pictures and compose music, and it still looks like a funny machine learning trick. But this is just a prelude. It will end when neural networks begin to learn how people take risks, make choices and how to use morality. Such studies are already underway, and we create data for training ourselves, often without even knowing about it. Naked Science is trying to imagine why neural networks are needed that mimic our manner of reasoning and behavior, who will benefit from this and what should be feared.
The program, created by researchers from the University of Toronto, Cornell University and Microsoft Research, guesses from a set of chess games who played them. It calculates the author by moves and can distinguish him from the games of thousands of other chess players who regularly play on the popular Lichess server. In essence, it determines the player’s inherent decision-making style.
Chess fans have long known that grandmasters have their own recognizable game style. Someone plays aggressively and is not afraid to take risks, while someone is cautious and waits for the opponent’s mistakes. There are those who are strong in openings, and others, on the contrary, are especially dangerous in endgames, when there are already few pieces on the board. In short, every chess player is unique, and there is something in his choice of moves that distinguishes him from all others. It is as unique as a fingerprint, a kind of “fingerprint” of style.
The program catches it, but it doesn’t care who makes the moves, the master or the novice amateur. She easily recognizes everyone.
The decisive contribution here is for machine learning — the authors took recordings of the games of players who played on Lichess at least a thousand times, and selected sequences of up to 32 moves from these games. They encoded each move — a change of position— in the form of numbers and transmitted it to a neural network, which represented any game as a point in a multidimensional space. For a neural network, all chess player’s games are a cluster of points (or a cluster). She was taught to maximize the cluster density of each player and the distance between clusters of different players.
This is how the neural network has learned to distinguish people — by how the moves of their parties converge into clusters. This cluster is the individual style of a particular player, which is not always expressed explicitly, but the machine sees it. Moreover, it distinguishes players with a high rating, even if it is trained only at amateur parties, and vice versa. The program really catches the individuality.
The authors of the study believe that the same can be done with poker. Or, they say, with the right data, such a program could identify people by the manner of driving a car or the time and place of using a mobile phone.
In short, instead of a set of moves in chess, there can be any digital traces. Any sufficiently long (digitized) history of behavior potentially contains data for training such programs. A person is recognized by characteristic chains of actions simply due to the fact that we are different, and each of us, even in small things, is somehow different from others. And if earlier, in order to confuse the tracks, it was possible to try to distort the handwriting or voice, then it is much more difficult to change the style of decision-making, it’s like replacing your psyche. Moreover, it is not known in advance what signs the network highlights and what exactly needs to be masked.
From style search to prediction: games and modeling of people
The authors of the program are concerned that their approach is suitable not only for chess, but also easily transferred to other areas: a neural network can be trained on any available data, and everything is not smooth with ethics. After all, not only scammers, but also ordinary people often want to remain anonymous, and not necessarily with malicious intent. Machine learning will make them visible.
In theory, this means that logging into the network under someone else’s IP will no longer help — any person can be calculated by his unique style, no matter what he expresses.
However, this is still only in theory. In practice, everything is not so simple: to train artificial intelligence, you first need to collect marked-up data, that is, separately record digital traces of each of the many millions of people present on the Internet, and preferably within months. And then constantly monitor them over the network. This requires serious computing power, and they, in turn, require additional energy. Finally, people walk on different sites, and their stories can be linked only if these sites actively exchange data with each other, which is hardly feasible (except for a number of exceptions — sites belonging to large corporations, like Instagram and Facebook, for example).
Large sites with a huge audience can provide themselves with such data. They will mainly collect digital traces of visitors, but primarily not in order to reveal their identities: it is much more promising to use these traces to study and predict behavior (for example, for marketing reasons). Social networks are a suitable testing ground for this. But the best one is massive online games.
The game effectively reveals the properties of the psyche. During the game, people make a lot of decisions and interact with other players in a complex and rapidly changing situation. They have to think tactically and strategically. They have to learn and gain experience. People have been playing some games for years, which means they accumulate a rich history of their actions. More importantly, millions of users play such games. All this creates huge amounts of statistics, it is more than enough for machine learning.
Those who started playing as a teenager can continue many years later, developing along with the game universe. During this time, their unique decision-making style will be deeply studied and determined, and such information can later become very valuable. Sometimes former teenagers become business leaders, high-ranking officials, politicians, military leaders of high rank. A machine trained on a large array of data will not just know in what manner they think and act, it will help to make predictions about them.
Of course, the accuracy of the forecast also depends on whether people will keep their decision-making style at a distance of decades. It is difficult to answer this question unequivocally, but longitudinal studies show that the main personality traits are quite stable from youth to adulthood. If a young girl is prone to reflection, she will dig into herself and in old age. If a young man is overly impressionable, then he will not lose this property with age.
Nuances can be smoothed out or developed, but the core of the psyche is difficult to change. You can reasonably bet that the peculiarities of thinking and perception, as well as temperament, people will carry with them all their lives. And if programs learn to catch these features, it promises profound consequences.
After all, the power of neural networks is not only that they find hidden patterns in a data set, they can still reproduce these patterns. The chess program created at the University of Toronto is able to play the way people play, predict the moves of a particular chess player and even anticipate typical mistakes that he will make in the game. She knows what mistakes players make at different skill levels, and can indicate the level at which people stop making them.
In other words, the program is not looking for the best move for a given position — it offers moves that a person would make. It models the decision-making process of chess players. This is forecasting.
From predictions to influence: Machines as psychologists
One should not hope that the matter will be limited to the artificial environment of chess. Last year, psychologists from Princeton published an article in the journal Science, “Using large-scale experiments and machine learning to discover theories of human decision-making.” The authors trained a neural network on a large database collected by various scientists over many years. It contains the results of psychological experiments on how people make risky choices, including gambling — in total, more than 10,000 different situations in which the subjects made certain decisions.
It turned out that trained neural networks are able to simulate human decisions with high accuracy, and they significantly exceed the models of risky choice previously proposed in psychology.
So machine learning has helped psychologists to create a new, more effective theory of behavior, which previously could not be developed. And this is not surprising: in trying to explain the choice of people, experts put forward hypotheses and rely on their intuition, but it is limited to experiments that the human mind is able to cover. No psychologist is able to dig through a huge database containing solutions of hundreds of thousands of participants in thousands of different choice situations.
For artificial intelligence, this will not be difficult.
What about moral issues? The Moral Machine project has already collected about 40 million solutions from people from more than 200 countries. This is the largest online experiment on moral dilemmas ever conducted. Participants are asked to decide in a traffic situation where an unmanned vehicle can turn in one direction or another. The subject must decide who to save and who to sacrifice. There may be different characters in the picture (for example, a man, a child, a female doctor, a dog) and different environmental options. Millions of unique tasks for moral choice.
Such a multidimensional space of solutions is beyond the power of the human mind. But a neural network trained on this data allowed psychologists to build “an informative, interpretable psychological theory that defines a set of moral principles underlying people’s judgments.” They write that this theory surpasses those that were invented earlier, and thanks to it they have identified three new effects.