MIPT has installed the world’s first computer for working with artificial neural networks. It will be mainly used by the Neural Networks and Deep Learning Lab for the iPavlov conversational artificial intelligence project. The head of the laboratory, Mikhail Burtsev, talks about the project, a newly emerging communication environment, and the importance of processing power for modern research.
Over the past few years, the number of messaging app users has far exceeded the number of social media users. Mobile messaging apps are used way more often than any other applications. Messaging has become the most popular activity on all devices. “We are essentially witnessing the rise of a new communication environment,” Mikhail Burtsev said. “This creates new opportunities for businesses, allowing them to build their communication with customers in a novel way. Users can now find all the information they need and get the answers to their questions without leaving the chat. Imagine that you won’t have to browse the internet in search of the most affordable plane tickets: All you have to do is text a bot, specifying the date and the destination of your flight, and it will suggest the best option. Messaging apps have the potential to become a channel for communication with companies and service centers: You will be able to text a bank, a store around the corner, a car dealership, etc., and get the service or consultation you need.”
Dialogue systems as such have been around for a long time, but the technologies for their development have been steadily evolving and improving. Such systems used to rely on preset answer rules, but when it comes to communicating with real people, this approach does not yield good results. The reason is we tend to convey the same information in many different ways, making it impossible for the computer to understand us. Right now, artificial neural networks are actually booming, and they perform particularly well in the context of natural language. For instance, switching Google Translate to neural network algorithms has significantly improved the quality of translation between languages. That is why the development of so-called conversational computer intelligence based on working with natural language data currently seems to be one of the most promising and exciting fields.
The iPavlov project, named after the pioneering researcher of higher nervous activity Ivan Pavlov, is focused on developing conversational artificial intelligence that would be capable of engaging in a meaningful conversation with humans. The research is carried out in line with the Neuronet roadmap of the Russian government’s National Technology Initiative, with the NTI project office — Russian Venture Company — providing organizational, technical, and financial support, along with expert analytics. The dialogue system developed within the project will be able to answer a conversation partner’s questions and request information needed to solve the task set in the dialogue. To do so, the neural network will learn from large arrays of documents and human dialogue transcripts.
The project aims to create a universal solution. The algorithm will essentially be able to learn to communicate in any language. This will be possible due to the interface of the system operating on a character-by-character basis, enabling training on data from any language regardless of its morphological structure. There are plans to use the new technology to develop a custom technological solution for the project’s partner Sberbank. The network will also be learning from real dialogues between bank staff and customers, which can eventually lead to creating a platform capable of communicating with customers on its own.
Photo. Researchers from the Neural Networks and Deep Learning Lab, left to right: Olga Kayrova, Rafael Ayrapetyan, Mikhail Burtsev, and Vadim Polulyakh
There is an aspect that sets work on AI apart from other previously developed advanced technologies: Most basic machine learning algorithms are now published under an open license. This is exactly what’s allowed artificial intelligence to become one of the most promising and rapidly advancing fields. “Possibly the best example of this is Elon Musk’s OpenAI project. Google, too, has open-source libraries — for example, TensorFlow — as well as thousands of other projects with open licenses. Such libraries are what sets the technological standards in the field of machine learning,” said Burtsev, whose laboratory also embraces this philosophy of open AI. As the final result of the project, the team plans to create an open platform with basic tools of conversational AI that will serve as the foundation for startups and new solutions in all kinds of fields.
The laboratory is planning to conduct its research on the world’s first supercomputer designed specifically for training artificial neural networks, a task for which computing power is vital. The better your hardware performance is, the more complex are the architectures of neural networks that you can work with. The complexity of the model can often spur a revolutionary leap in solving practical tasks. For instance, the current revolution in computer vision and speech recognition is, among other things, associated with a surge in computing power. “The power of two computing nodes of the Nvidia DGX-1 system is enough to make it one of Russia’s top 50 high-performance computers. We will use DGX to experiment with complex models of dialogue agents. This will allow us to solve practical problems we would not have otherwise been able to tackle,” Burtsev explained.
The nodes of the supercomputer are installed in the MIPT Data Processing Center, so other research teams at the University that work with artificial neural networks can also use it to conduct their research.
The original story was written by Anastasiia Grachikova.