Research Professor in the Faculty of Computer Science’s Big Data and Information Retrieval School
Forecasting is an unforgiving endeavour, and this is particularly true as progress in the field of contemporary machine learning methods has been surpassing even the boldest of expectations over the last several years. But I’ll venture to name a few areas in the study of deep neural networks that can be expected to make great advancement in the near future.
First is the development of ideas with reinforcement that will allow for new machine learning algorithms to be developed for agents to interact with the environment. This might be a robot or virtual reality programmes for playing intellectual games like Go (already made) or StarCraft (in progress). The main goal here will, of course, be to create an algorithm that can adapt to a new complex game or environment ‘on the go.’
Second is the development of new ‘on-the-fly’ learning methods and meta-learning. The first allows the computer to grasp new concepts and ideas using several examples similar to how a person does. This is not like contemporary neural networks that learn a new concept after seeing thousands or hundreds of thousands of examples. The second, meta-learning, allows a neural network to select the parameters for learning methods itself. Currently, the quality and speed of neural network training depends heavily on how certain parameters are aligned (usually so-called hyper-parameters in order to differentiate between these parameters and the different weights the network establishes during learning), and these factors also depend on the architecture of the network itself. These are currently determined either by a person or by semi-automatic procedures that are far from optimal. Because of this, neural networks learn more poorly and for a longer amount of time than they could.
Studies from 2016 show that this work can largely be entrusted to an auxiliary neural network. As we all remember from middle school, a sign that an industrial revolution is over occurs when ‘machines start producing machines.’ It is possible that, in the future, an equally important milestone will be when neural networks begin teaching neural networks, and there is reason to believe that this will take place in as early as 2017.
Third, neural networks will learn to talk to humans (both in the sense of replicating texts, and in the sense of synthesising speech indistinguishable from that of humans), generate photorealistic images and video footage based on a text description, and write large and meaningful texts. This will be what our near future looks like thanks to the considerable progress being made in the field of generative deep learning models. All of this will, of course, lead to the creation of new businesses; the emergence of new types of goods and services; and to the growth in workforce productivity in traditional sectors of the economy, such as cellular operators or banks, which will be able to do away with expensive and ineffective call centres.
Solving these tasks will be an important step on the road towards the Holy Grail of machine learning — the creation of artificial intelligence. AI will happen next year, of course, but it will unquestionably be developed in the next 5-10 years. In addition, the elements of AI that already exist will help scientists create full-fledged AI and also allow for work in this field to take place at a much faster pace. Creating AI will be one of humanity’s most important achievements, and it will guarantee a powerful leap forward for human civilization.
It is important to note that the rapid progress seen in the field of artificial intelligence was possible largely because these developments take place openly; any person who has the minimum necessary training — a graduate of the HSE Faculty of Computer Science, for example — is able to take part. Even top IT corporations have no secrets when it comes to deep learning, aside from short-term commercial ones. In addition, software for the majority of methodologies is openly available, as are mathematical descriptions of algorithms that are surprisingly not as complex as one might think given the universality of the problems these algorithms solve. This distinguishes machine learning from, say, space or nuclear programmes of the mid-20th century.
MIT textbook on Deep Learning
On the eve of New Year’s, it is customary to take a look into the near future. We asked HSE experts in various fields to share their forecasts on which areas of research might be the most interesting and promising in 2017. They tell us about what discoveries and breakthroughs await us in 2017, as well as how this could even change our lives.