“I am quite speechless”, said Lee Sedol, one of the best Go players of all time, when AlphaGo, artificial intelligence (AI) created by Google, beat him in the five-game match in 2016. Even though already two decades ago IBM computer beat world chess champion, until recently the ancient Chinese board game Go has been considered as a much more difficult game for an artificial mind to master. Due to high number of alternative moves it has been claimed that strategy is not enough to win the game and that victory required intuitive intelligence. Nevertheless, by applying a machine learning technique known as deep learning which makes computers to learn multiple levels of abstraction and representation, AlphaGo analysed thousands of games and developed its intuitive sense of strategy which secured a victory of an artificial intelligence over a human mind.
The victory of AlphaGo symbolises a progress which has been made in the last decade. AI machines are already superior to humans when dealing with vast amount of data and this competency found applications in many fields and industries, starting with agriculture and travelling and finishing with business, law and medicine. Nonetheless, while the achievements in the field of AI are breath taking machine intelligence is still quite far from human intelligence. Understanding language, ability to learn and to reason is what makes us, humans, smarter than AI. At least for a little while.
Understanding figurative language
Today to process and understand texts, machines must convert natural language into data. However, so far machines do not have a contextual understanding of the text they are processing. The problem is that in any language same words might mean various things in diverse contexts and different words might represent similar meanings. For example, a request to interpret a well-known idiom – “Money does not grow on trees” – would expose any artificially intelligent machine to an unsolvable problem. While each single word would be correctly understood, a deeper meaning – that money is not that easy to come by – would not be captured by a machine unless the exact meaning of the idiom would be in advance provided to the machine.
To teach AI to understand figurative speech scientists come up with idea to represent words as mathematical vectors which would allow similarities between words to be estimated in this way teaching machines to sense a context of a sentence and catch meaning hidden behind words. For instance, even though two words- “sky” and “plane” – do not look similar, they are close in vector space. The idea is now used to build networks in which each word in a sentence can be used to construct a more complex representation and what Geoffrey Hinton, a deep learning researcher at Google, calls a “thought vector”.
Learning by observing
Another mountain that AI developers need to climb is teaching machines to observe. Apparently, most of what humans learn is learned during the first few years of life by observing and interacting with the world. For example, a two-year-old understands physics without taking classes about gravity and pressure – by only playing and observing. However, scientists today put much more effort to make machines learn. AI machines use algorithms to learn mapping function from the input to the output. It is called supervised learning because the process of an algorithm learning from the training dataset can be viewed as a teacher supervising the learning process. Basically, for a machine to learn to recognize an elephant thousands of different images of elephants must be provided to the machine. Obviously, this learning process requires a vast amount of data to be collected which is inefficient. In addition to that, supervised learning does not train machines to predict under uncertainty and learn to take what has been learned from one task and apply it to another.
To address challenges imposed by supervised learning two relatively novel solutions can be mentioned. First of them is a technology presented couple of months ago by a Boston-based startup called Gamalon. The technology builds algorithms capable of learning from fewer examples. It employs probabilistic programming—or code that deals in probabilities rather than specific variables—to build a predictive model that explains a particular data set. In turn, from just a few examples, a program can assume, for instance, that it’s highly probable that elephants have big ears, tails and long trunks. The more examples are provided, the more times code behind the model is rewritten, and the probabilities increased.
Another solution to address the drawbacks of supervised learning is AI technique called “generative adversarial networks” (GANs). The technique is being developed by Ian Goodfellow, one of the world’s most important AI researchers who works at the Google Brain lab. The idea behind the concept is to create two interrelated machines which can learn from each other. For example, one AI creates realistic images, while a second AI analyses the results and tries to decide whether the images are real or fake.
Both probabilistic programming and GANs might allow to make a big step towards unsupervised learning which is like learning by observing the world. Unsupervised learning would decrease the amount of human labour needed for training machines and maybe even would lead to the point when machines will be able to build their own learning models without any help from humans.
Ability to reason
The more data is absorbed by AI and the tougher tasks are given to it, the more difficult it is to understand an explanation for the AI provided solution. In the case of earlier mentioned Go game, reasoning of certain moves made by AI is interesting and yet not critically important. However, visibility of decisions is of extreme importance when it comes to adopting AI in such fields like healthcare or self-driving cars where we cannot blindly trust a machine due to arising accountability issues. And trust is possible only if we can ask a machine “Why?” when its answer is not intuitive to us and receive a clearly presented explanation in return. Even so, it is expected that developments in linguistic AI skills might one day help to teach machines to reason in an understandable and trustworthy manner.
Beyond the doubt, AI is much better at calculating than humans but it seems that we can sleep tight because today AI is still way behind a human mind and despite many theories currently available it might require many years to see their practical implications. For instance, it took 40 years for the deep learning technique to prove its potential in the development of AI. On the other hand, once it has been proven less than couple of years (instead of decades estimated) were needed for the developers of AlphaGo to program a machine capable of winning over a human being. Well, at least today neither humans nor machines can estimate the future, therefore we will just have to live and see how quickly AI will catch up with a human mind.
Jana Sadauskaite