Experts comments Artificial Intelligence

Top-10 technology trends AI in 2018 year

Artificial intelligence (AI) is now under the watchful eye of the public, leading business leaders and government agencies around the world. We live in an era of innovative technologies that determine the future. But what is happening behind the scenes of controversy - in the world's laboratories and research centers, which in fact set the course of AI development for the next decade? Leading industry experts have learned the latest developments of technology, identified key trends and explained why they are important.

1 The theory of deep learning: demystification of the work of neural networks (CNN)

What it is?

CNN (convolutional neural network) - deep neural networks whose application theory appeared in the 50-ies, but no one understood how to fully implement machine learning, and people began to surrender. However, the desired results were still achieved. CNN is working, accumulating many functions at each level. Starting with the search for edges, then the figures, and then the actual objects. However, previously information about the spatial relationship of all functions was lost, and this forced the scientists to apply new theories and schemes of work with a digital array. Today, deep neural networks demonstrate their ability to "learn", receiving information not only from images, but also from text and audio data. Every year, artificial intelligence becomes smarter, as new technologies allow you to apply methods of deep machine learning. As a result, today neural networks can not simply analyze and organize information, but also forget or compress data that is less important for certain purposes, and highlight the ones that are most important for achieving a holistic result.

Why is this important?

An accurate understanding of how deep learning works, contributes to its wider development and application. For example, the optimal choice of design and network architecture can be made more obvious, while providing greater transparency for high-reliability systems or control applications. In the near future, more results are expected from the study of this theory in its application to other types of deep neural networks and its development in general.

2 Capsule Network (CN): Simulation of visual processing of the brain

What it is?

CN (Capsule Networks) is a relatively new type of deep neural network that processes visual information in almost the same way as a brain, which means the ability to maintain hierarchical relationships. The new architecture provides improved accuracy in each subsequent set of data. This set was carefully designed to become a complete form recognition task. If we speak in detail, then in this approach are used small groups of neurons - capsules, which form layers for identifying objects on video or images. When several capsules in the same layer take the same decision, they activate another capsule that is at a higher level. This process continues until the network is able to conclude that it sees. Each of these capsules is designed in such a way that it detects certain features in the image and recognizes it in different scenarios. This is about the ability to recognize objects in general, even if they are inverted or shown at a different angle.

Why is this important?

CN has reduced the number of standard CNN errors by 45%. Due to the new approach, its networks require less data to recognize objects in new situations. The published reports show that capsular networks are not far behind ordinary when it comes to identifying, for example, handwritten characters. They also make fewer mistakes when trying to recognize previously seen objects at different angles. Of course, it's too early to talk about whether such systems can become an alternative to traditional neural networks, but one can expect that machine-learning enthusiasts will implement this approach in practice and find the answer to that question.

3 Deep Reinforcement Training (DRL): Environmental Engagement for Business Tasks

What it is?

DRL (Deep reinforcement learning) - the type of neural networks that learn by interacting with the environment through observations, actions and rewards. Deep reinforcement training was used to study gaming strategies such as Atari and Go, including the famous AlphaGo program that won the man.

Why is this important?

DRL is, in fact, the most versatile of all learning methods, so it can be used in most business applications. It requires less data than other methods for learning their models. In addition, it can be taught by simulation, which completely eliminates the need for data labeling. This is a great advantage, so we're likely to see new business applications that combine DRL and agent simulation already this year.

4 Generational Advent Network (GAN): Combining neural networks to stimulate learning and facilitate computational stress.

What it is?

GAN (generative adversarial network) is a type of deep learning system that does not require a "teacher". It is implemented in the form of two competing neural networks. One network generator - creates fake data that looks exactly the same as a real data set. The second network - the discriminator - processes the real and already generated data. Subsequently, each network is improved, allowing the couple to study the entire distribution of this data set.

Why is this important?

GANs offer a wide range of opportunities for solving learning problems without a teacher, in which the labeled data does not exist or is too expensive to receive. They also reduce the load required to implement a deep neural network. It is expected to see more business applications, such as cyberattack detection, using GAN.

5 Training on incomplete (LD) and supplementary data: solving tasks with marked data

What it is?

One of the problems of machine learning is the availability of large volumes of labeled data required for the learning system. Two methods can help solve this problem: synthesizing new data (1) and transferring a model prepared for one task or region to another (2). Methods such as the transfer of learning (transfer of knowledge from one task / area to another), or one-time training (transfer of learning that occurs only with one or without suitable examples) - and there is a technique of so-called LD (Lean Data), training on incomplete data. Similarly, the synthesis of new data through simulation or interpolation helps to obtain more data, thereby complementing existing data to improve learning.

Why is this important?

By using these methods, we can solve a variety of problems, especially those that do not have integral input. It is anticipated that we will soon see more options for incomplete and supplemented data, as well as different types of training that are used to address a wide range of business tasks.

6 Probable programming: "languages" to facilitate model development

What it is?

This is a high-level programming language, which greatly facilitates the development of probable models, and then automatically "solves" these models. Probable programming languages ​​allow you to reuse model libraries, maintain interactive simulation and formal verification, and provide the level of abstraction needed to create a common and effective output in universal class models.

Why is this important?

Probable programming languages ​​have the ability to take into account uncertain and incomplete information that is commonly distributed in the business area. In the future, we will see a wider introduction of these languages. It is expected that they will also be applied to in-depth training.

7 Models of hybrid learning: a combination of approaches to model uncertainty

What it is?

Different types of deep neural networks such as GAN or DRL have shown great prospects in terms of their performance and wide application to different types of data. However, models of deep learning do not simulate uncertainty, as do "Baisovsky" and probable approaches. Hybrid learning models combine two approaches - in order to use the strengths of each of them. Some examples of hybrid models are "Baeysovskaya" deep study, "Baeysk" GAN and "Baisovsky" conditional GAN.

Why is this important?

Hybrid learning models can expand the variety of business tasks, including deep learning with uncertainty. It can help us achieve better performance and understanding of models, which, in turn, will facilitate their more extensive implementation. It is expected that deeper methods of learning will receive "Baeysk" equivalents, and a combination of probable programming languages ​​will begin to include profound learning.

8 Automatic machine learning (AutoML): Creating a model without programming

What it is?

Machine Learning for Automated Algorithm Design (AutoML) - can automate the workflow of developing learning models. In this case, a number of different methods of statistical and in-depth training are used, since in the beginning this process is very complex and requires observation by experts (data preparation, choice of functions, choice of model or technique, training and setting).

Why is this important?

AutoML is part of what is considered as democratization of AI tools, enabling business users to develop machine learning models without deep programming. It will also speed up the time spent by scientists to create models. The next year, it is expected to see more commercial packages of AutoML and the integration of AutoML on significantly larger machine learning platforms.

9 Digital Double: Virtual Replicas Outside of Industrial Applications

What it is?

Digital Double is a virtual model used to facilitate detailed analysis and monitoring of physical or psychological systems. The concept of a digital twin emerged in the world of production, which was widely used for the analysis and monitoring of things like wind farms or integrated industrial systems. Now, using agent-based modeling (computational models for modeling actions and interaction of autonomous agents) and system dynamics (a computer-based approach to the analysis and modeling of behavior lines), digital twins are applied to non-physical objects and processes, including prediction of consumer behavior or buyer

Why is this important?

Digital twins can promote the development and wider adoption of the Internet of Things (IoT), providing a way of predictive diagnosis and support for IoT systems. In the future, wider use of digital twins is expected in both physical systems and consumer choice modeling.

10 Clear Artificial Intelligence (AI): Involving the Black Box Method

What it is?

Today, there are many algorithms of machine learning that can actually feel, think and act through various applications. However, to this day, some of these algorithms are actually "black boxes", shedding too little light on how they achieved such results. Understanding AI is another step towards new methods of machine learning that create more understandable models while maintaining the accuracy of prediction.

Why is this important?

AI - Clear, Evidence-based and Transparent - will be critical to building confidence in this technology and will contribute to a wider introduction of machine learning. Firms will apply a clear AI as a requirement or best practice before embarking on widespread deployment of artificial intelligence technologies, while governments will be able to make the AI ​​understandable in the future.

everest.ua