Machine Learning: Potential and Opportunity



    / photo Jaro Larnos CC

    Cloud technologies are widely used in various scientific fields: once we talked about how clouds are used in physics and astronomy , as well as geography and genetics . Virtual infrastructures allow scientists to process a huge amount of information in the shortest possible time, which leads to new discoveries.

    But there is another technology that can change our understanding of information processing. We are talking about machine learning, which has recently gained particular popularity.

    A few years ago, Google completely redesigned its image search and speech recognition services by introducing machine learning elements, and on June 16 this year, a giantannounced the expansion of a research center in Zurich, which will be engaged in developments in the field of AI, natural language processing and machine perception. This means that Google is going to develop systems that are able to think, listen and see.

    Greg Corrado, a senior fellow at the company , argues that actively introducing machine learning can be as useful as creating the Internet. This can lead to the fact that we will not need to understand all the details of certain processes, it will be enough to upload data to the system, based on which it will begin to learn.

    The most promising area in machine learning is the so-called deep learning. It is built on neural networks (NS), which require a large amount of data to learn. NSs were first described in the 30s of the last century, but they began to be actively used only in the last 3-4 years, as the power of computers increased sharply.

    Last year, Google made public its TensorFlow deep machine learning library . So the company is trying to draw attention to the project and develop it with the help of third-party developers. The system presents the calculations in the form of a data flow graph. Its main feature - unlike other platforms like Theano and Torch - support for distributed computing.

    TensorFlow uses almost in all directions: from speech recognition to photo search, but in reality it will be more useful to scientists conducting experiments on deep training of neural networks, as well as companies that need to quickly train and test their models. You can feel TensorFlow yourself by clicking on this link.

    AI goes to writers


    The Guardian journalist Alex Hern in his article spoke about his attempt to train the simplest recurrent NS so that she could logically complete the proposals. As training data, he took 119 MB of text from The Guardian articles. Read about other interesting applications of recurrent NS in this article.

    Half an hour after starting the training process, Alex saw that progress was only 1%. He realized that the power of his computer was not enough and decided to rent a server in the cloud. This allowed us to complete the learning process in 8 hours.

    It turned out, to put it mildly, not very cool. The computer needed to continue the following sentence: "The fateful decision to stay in the EU, made on Thursday, was ...". As a result, the system proposed options such as "... based on a promise made in several statements" and "... a member of the 2015 opposition party." On the one hand, complete nonsense, on the other - there is also a positive point: if the machine learned to write articles for The Guardian, Alex and his colleagues would be left without work.

    This result is quite explainable. The neural network used in training could recognize only symbols: she did not know what a word was and did not understand grammar. Moreover, in order for the network to be able to adequately draw up proposals based on data about the real world, it needs to transfer a much larger amount of data for training. A set of articles from one publication is not enough. Journalist The Atlantic and a writer who previously worked on Twitter also wrote about such experiments .

    Humanity rushes to the rescue


    One of the most striking examples of deep learning is AlphaGo, an AI-based program that has recently beaten the go world champion. The program involves two types of training: training with a teacher, when using the data of all matches played between people, and training with reinforcement, which implies that the program plays against itself and learns from its mistakes. But with all this, as it turned out, AlphaGo just can’t learn some things on its own.

    According tothe leader of the program’s development team DeepMind, the system understood well what areas of the playing field it should focus on. However, the program does not know when it should stop the “thought process” and make its move. This is an important point in the game, as in professional matches there is a complex system of time control. For example, in the game against world champion Lee Sedol, the players had two hours to think about the moves, as well as three minutes of extra time, which were added if none of the previous moves were completed in less than three minutes.

    The developers did not add time control rules to the program, but only introduced a restriction by developing a special algorithm. Later it was optimized by the program on the basis of a series of experiments, but the fact is that without human help AlphaGo could not have beaten the champion.

    This situation with AlphaGo leads us to think that the progress of AI training can be accelerated if ordinary users are involved in learning systems. For example, a popular computer game Minecraft is now becoming a platform for collaboration between humans and cars.



    Recently postedat the GitHub Project, Malmo, launched by Microsoft, is a platform for exploring the capabilities of artificial intelligence. The task is to train the character of the game to perform various actions, starting with crossing the bridge and ending with the construction of complex objects. In addition, the project allows you to organize a joint AI game with a person, as well as communication between them using a special chat.

    According toProject Manager Katja Hofmann, Project Malmo's goal is to create an AI that will learn from users and help them solve their problems. The program uses reinforcement learning algorithms. For example, you can teach a car to navigate in a room with many obstacles. Ordinary players can give hints or instructions that AI gradually learn to recognize and make the right decisions based on them.



    The Minecraft platform was also used to train the robot at Brown University. According to one of the university professors, Project Malmo will become an effective method of collecting data on human interaction with AI. Perhaps in the near future we will be able to fully communicate with artificial intelligence.

    PS Additional Machine Learning Resources

    • Machine Learning Courses from Coursera and Udacity ;
    • Udacity Deep Learning Course ;
    • Video lectures by Professor K.V. Vorontsov in machine learning;
    • TensorFlow Mini Tutorials for Beginners
    • Useful post by graduate student Andrei Karpathy from Stanford on the use of recurrent neural networks.

    PPS And here are some more interesting materials from our blog:


    Also popular now: