Trends in neural networks - innovations and forecasts

Nowadays, the use of artificial intelligence is fashionable and a trend in itself. AI optimizes costs, reduces risks, helps conduct scientific research, generates creative ideas, monitors security and entertains. In this article, we will talk about trends in neural networks, the development of new architectures, the possibilities of integrating AI with other technologies and forecasts for the development of artificial intelligence in the near future.

Latest architectural solutions

In 2024, developers presented an alternative multilayer perceptron (MLP), called Kolmogorov-Arnold Networks (KAN) in honor of Andrey Kolmogorov and Vladimir Arnold, famous Soviet mathematicians.

The fundamental feature of KAN is moving the activation function not to the core, but to the edges of the neuron. This accelerates the learning rate and increases interpretability. Due to nonlinear data processing, KAN eliminates “catastrophic forgetting” in machine learning. When a neural network, receiving new information, loses the ability to process old information. Since when new data is received at the input, the weights are redistributed and the values are rewritten.

In addition, Kolmogorov-Arnold Networks require fewer inputs, it is easier to visualize for easy interaction with users. Plus, AI shows higher accuracy compared to MLP. According to forecasts, the architecture will open up new opportunities for the development of AI.

In 2020, the developer of Liquid AI presented the liquid neural network LNN, which can learn in real time, while “remembering” previously received information. But the main advantage of LNN is the minimum number of neurons. For example, an algorithm with a standard architecture needs at least 100 thousand neurons and up to half a million parameters to keep a car in its lane.

A liquid neural network needs 19 neurons to keep a car on the road, assess the road surface and the turning angle. During the experiment, the neural network automatically reduced speed on sharp turns and dirt roads.

Application in various industries

Systems based on artificial intelligence have already been implemented in many economic sectors.

In construction - computer vision technology searches for defects in concrete, metal, processing images in real time and transmitting photo reports to workers. Monitors compliance with labor safety requirements. In Japan, for example, algorithms have been trained to determine the risk of workers falling from scaffolding. Sensors track the movement of builders, and AI identifies patterns and warns of danger.
In industry - predict possible equipment failure, reducing companies’ costs for major repairs and replacement. Determine the presence of dangerous objects in scrap metal, replace quality control, identify defects and independently generate electronic passports.
In marketing - personalize advertising, conduct competitor analysis, update the semantic core for text content. Manage mailings, prepare reports on the effectiveness of advertising campaigns.
In sales – AI predicts demand, sales dynamics, assesses risks, takes on routine processes. For example, answering phone calls, filling out customer cards, tracking the status of a transaction.
In medicine – AI algorithms fill out documents, analyze X-rays, identifying common symptoms. Which allows identifying diseases at early stages. In pharmacology – the machine creates drug formulas, makes forecasts on how components will interact with each other in theory.
In education – helps teachers create presentations, generates text and pictures. A neural network can explain a complex topic to students (for example, on programming) in simple language, translate the text. Compress a difficult to perceive scientific article into a short summary containing the main facts.
In trading – they form an investment portfolio, analyze charts, trading statistics, quotes to assess the profitability of investments. They predict the growth and fall of asset prices taking into account economic and political factors.
In agriculture – algorithms monitor the health of animals and birds, calculate the optimal time for sowing. They help assess the level of ripeness of fruits and vegetables.
In the extractive industry – creating digital prototypes of rocks to determine the ratio of gas and oil, searching for deposits.
In transport – developing logistics solutions, including during the construction of enterprises, to eliminate extreme loads on equipment or downtime, crossing routes.

Cameras with computer vision technology monitor order in crowded places, record violations on roads. Help detain criminals.

Integration with other technologies

The ability of AI to quickly and accurately process large amounts of data, the speed of model training made it possible to integrate neural networks with:

CRM – a combination of machine learning, natural language processing, in addition to automating routine business processes, significantly improves communication with clients;
monitoring systems – road networks, equipment, personnel increases safety, reduces costs, allows you to evaluate the effectiveness of employees;
applications for processing photos, videos – to improve quality, increase resolution, remove foreign objects, eliminate noise;
services for web design – font selection, image generation based on description, creation of 3D animation, objects, textures in virtual space;
medical devices – based on the patient’s medical history, generate treatment protocols, offer options for additional examinations. By the way, the last point refers to software. According to FDA standards, software that uses statistical indicators to compare common symptoms is not recognized as medical devices.

In 2024, Yandex conducted research, according to the results of which 20% of the largest Russian companies use neural networks in their work. These are retail, banks, industrial, mining corporations. Of these, 66% used AI in marketing, sales, 54% to improve customer service and 49% in research and development. But businesses have already appreciated the prospects of digitalization, so the numbers will only grow every year.

Forecasts

Digital development in Russia is a long-term government project. By 2030, it is planned to introduce AI services in all sectors, from medicine, education, to agriculture and industry. Develop measures to organize the security and transparency of AI-based models.

Software with AI is already used in the work of:

Magnit - predictive analytics;
Sber – customer scoring, chatbots, voice assistants, debt analysis;
T-Bank – created virtual investment, financial, travel, shopping assistants;
Alfa-Bank – personalized mortgage loan calculation, without loan registration;
Vkus-Vill – packaging design generation;
Epica – advertising campaign conducted by neuroinfluencer Summer;
MegaFon – advertising with the image of Bruce Willis;
SDEK – communication with clients in chats.

Ozon is testing a development that should simplify the business of sellers. Sellers will be able to “try on” clothes and shoes on virtual models, generate pictures, videos and place them in the product card. This will reduce costs and time spent on photo sessions.

You can evaluate what AI models can do for free, without downloading applications - right on the site. Access will open immediately after registration. There is no need to apply for a foreign bank card or phone number. You can work in the browser extension, through bots in TG and VK or on the website.

Results

Neural networks will develop and find new applications. Already now, fundamentally different architectures are being developed that require less data and computing power.