The main component of modern artificial intelligence is deep neural networks that provide an efficient machine learning process. Neural networks allow to solve issues related to computer vision, analysis and processing of any data types. The technology is used everywhere: from robotics to online marketing. We talked to Dmitry Korobchenko, a machine learning engineer at NVIDIA, about the application of neural networks and the prospects of artificial intelligence.
Interviewer: AI Conference (AI)
Respondent: Dmitry Korobchenko (D.K.)
AI: What was the most attractive for you in learning operation features of neural networks? When and why did machine learning engage you?
D.K.: I started exploring machine learning at the faculty of Computational Mathematics and Cybernetics of Lomonosov Moscow State University about 10 years ago. At that time, I was interested in and impressed by the computer vision task. And, as it turned out, one had to apply machine learning to solve successfully such a problem. Then, neural networks in our environment were still considered to be something indecent and ridiculous after all of their failures.
Nevertheless, I was intrigued by the common concept of machine learning related to various predictions based on the training sample or the search for hidden data patterns. At the same time, computer vision seemed to be especially appealing because of combining two interesting areas: machine learning and image processing. At the beginning of my career at Samsung, I explored neural networks more detailed (prior to the boom of this issue).
Neural networks amazed me by their simplicity and efficiency for solving computer vision tasks. If you have enough data for learning and the possibility to train using GPU, you do not have to contrive and create complex features for recognition system operations but you should just build a convolutional neural network, which will do everything on its own.
AI: What is the main advantage of the neural network?
D.K.: Unlike complex, designed manually feature extraction methods, the convolutional neural network is just a composition of quite simple and user-friendly operations. So, its appropriate features appear automatically during training.
Other types of neural networks have similar advantages. One more motivation for exploring neural networks is a constantly growing flow of various upgrades and inventions of new architectures with the support of scientific and industrial community and the absence of any standstill.
AI: What successes in the global business development were achieved due to the application of AI technologies? What achievements can we see in the future?
D.K.: Every company (IT, production, service etc.) starting applying AI technology in its products has already obtained some benefits from it or has made a good groundwork for its further development. Modern AI allowed to solve a range of issues that previous generation algorithms failed to solve.
Here is a partial list of examples: competently improved speech recognition in voice assistants, upgraded machine translation, analysis of users’ comments in numerous services, analysis of camera images at manufacturing sites to determine abnormities and detect other significant things, face recognition on social media and in banks, users’ analysis, prediction of their behavior, robotics, autopilot, and other smart gadgets.
In the future, the AI integration into manufacturing and products/services will be a common thing, allowing to significantly reduce the amount of human routine work and enhance the quality of products and services.
AI: Besides well-known sectors applying AI technology, where else can it be integrated?
D.K.: Wherever there is data and the necessity to analyze and process them, for example, recognition, content synthesis, robot or agent control, transferring of one data type into another, decision making, predicting etc. The main challenge is not ideas but a real possibility to integrate AI into a certain industry.
Everything is frequently limited to the following problems: the lack of data for training, poorly formalized, and unstructured data that are difficult to use in any algorithm, or the absence of an appropriate and efficient neural network architecture for the given task (data type). Therefore, some ideas die without managing to grow and fail to become well-known.
AI: Tell us please about the computer vision theory. Why do people need “new eyes”?
D.K.: Computer vision theory is a bridge between the visual information world and the business intelligence world. With images and videos playing a significant role in the modern data sector, one should know the tool for processing such data type. Applying any obvious image properties, such as a picture size, color statistics and so on, is not enough for this purpose.
Frequently, one should analyze content in the semantic meaning: what is depicted on this image, which objects it includes and where they are located, what the relation between these object is etc. The same image analysis method is required for robot vision, allowing robots to see the world they interact with. Autopilot vehicles can be also referred here.
AI: What are the difficulties of the development of computer vision technology?
D.K.: The computer vision task is not easy. Digital images are huge numerical tables (intensity in pixels) where it is quite complicated to distinguish a certain object. To solve this problem, one requires an intermediate stage: feature extraction.
Here is an example: to find a cat on the photo, one should initially find its ears, eyes etc. Such visual features have been specialized and designed manually for a long time. The advent of convolutional neural networks in computer vision has provided an opportunity to train these features automatically using data.
AI: How does deep learning technology help to improve the efficiency of online marketing?
D.K.: It is significant to solve the following online marketing tasks: content personalization for various users, detection of users’ features, prediction of their behavior, clusterization, recommendations, and other things. All of this can be solved using modern machine learning methods, including neural networks.
Such tasks often involve complex data types (images, texts). Deep neural networks (convolutional, recurrent) are the most appropriate for intelligent processing of this data type. However, neural networks along with other machine learning algorithms can be successfully applied even in the analysis of low-dimensional feature vectors (for instance, the list of certain user’s characteristics or actions).
AI: Let’s daydream about ways of future development. What place can artificial intelligence occupy there according to the boldest imaginations?
D.K.: In the future, artificial intelligence will become a multipurpose common tool for solving various tasks related to any data types. Almost all of our current office work will be performed by AI. One can already automate a lot of routine things using AI. Artificial intelligence will be responsible for more and more intelligent tasks.
Doctors lawyers, teachers, managers, and even scientists and engineers will have to compete with AI. Creative professionals can also face the same destiny, because AI will learn how to create richer content: automatically synthesized movies, music, novels etc. Besides, along with robotics development, AI will be able to perform not only virtual tasks but also make a lot of things in the real world, carrying out operations earlier conducted only by the human being.
AI: How beneficial will be your lecture for people?
D.K.: The lecture will provide a wide review of modern neural networks, their types, operation and training concept, as well as case studies. Besides facts about how it works, the audience will be able to realize the core of these algorithms and understand why it works.
Such an understanding will allow to treat modern artificial intelligence technologies not as magic black boxes but as quite regular analysis and data processing algorithms.
AI: And the final question: can one feel like the God, developing artificial intelligence?
D.K.: What we call ‘artificial intelligence’ is the same algorithm, tool, and software as other computer science things. Therefore, an engineer creating such an algorithm can feel like the Creator if their algorithm functions seamlessly and makes something incredible.
This can be also referred to almost any efficient and creative activity. However, it is too early to talk about the development of strong artificial intelligence.
Thanks for the conversation.
Artificial intelligence is involved in almost all the business processes: marketing, manufacturing, customer assistance, and sales. Deep neural network technology is increasingly applied because of machine learning.
At AI Conference, Dmitry Korobchenko, NVIDIA engineer and an expert in this sector, will speak on the topic: Deep learning: technology review and case studies. He will reveal how modern neural networks operate and learn as well as explain what Deep Learning is and where it can be applied.
Attend the conference and discover possibilities and prospects of AI application in businesses.