How Neural Networks Took Over the Digital World
Attention to artificial intelligence (AI) has regularly increased following the release of the latest sci-fi techno-thriller. "Terminator," "The Matrix," "I, Robot," "2008: A Space Odyssey"—in the past, these films sparked a real excitement about "machines" and gave rise to numerous conspiracy theories. But in 2023, the talk was not about a movie or a new book, but about the simultaneous emergence of several neural networks with creative abilities—Chat GPT, Mid Journey, DALL-E.
Master Chat GPT and Enhance your skills in in-demand areas—28 courses on one platform for 49.90 BYN per month.
Learn moreArtificial intelligence (AI) has been around since the 1960s. These are mathematical systems and algorithms that simulate human cognitive abilities. Initially, such computer models were used to study the human brain and mind in general. But their scope is much broader. For example, AI includes algorithms that control non-player characters in computer games, collect data from the network according to a given template, control industrial robots, and analyze large data sets.
Artificial intelligence does not necessarily have to think and make independent decisions like humans do—we're not talking about self-awareness in principle. It imitates the activity of the human brain, performing routine tasks, but can also act according to a predetermined program. For example, this is how the supercomputer Deep Blue, which beat Garry Kasparov at chess, operated with specialized software.
AI is divided into three types based on its level of development:
- Artificial Narrow Intelligence, ANI — specialized AI.It is designed to perform one task. This includes Deep Blue, voice and face recognition systems, and chatbots (text models) that can write poetry, autopilots, and text translators.
- Artificial General Intelligence, AGI — general AI.Multitasking artificial intelligence, comparable in capabilities to humans. Such AI can write a poem, set lyrics to music, create a picture or video, and in its free time trade stocks and drive a car. It may have consciousness, like in the movies "Chappie" or "Her". Of course, such a level of development is not yet available and is at the research stage.
- Artificial Super Intelligence, ASI.Superintelligence, the cognitive abilities of which are much higher than those of humans. An example of such AI is well depicted in the movie "Transcendence".
One of the areas of artificial intelligence is machine learning technologies. While in conventional applications a database is connected in advance, which is used to execute the embedded algorithm, such systems learn on their own. For example, an algorithm is written, hundreds or thousands of images of license plates are loaded into it, and the equivalent (letter, number) is specified separately. The process is similar to how small children learn to read. After training, the system will be able to recognize license plates in images. Using this method, it is possible to teach AI to recognize any images—numbers, texts, animals, human faces, and speech.
Text recognition applications (Abbyy Scan, Abbyy FineReader), video surveillance cameras with face or license plate recognition (computer vision), autofocus and image enhancement systems in cameras and smartphone cameras, speech recognition and its conversion to text are based on machine learning technologies - all these are examples of AI. Similar technologies began to be used back in the late 90s.
Neural networks are a development of machine learning technology. These are systems that can not only recognize images and learn from them, but also apply the acquired knowledge and adapt to new conditions. In fact, this is a complex of programs that possess cognitive characteristics and capabilities inaccessible to conventional computers. First of all, this is the ability to learn from one's mistakes, identify hidden patterns, and demonstrate creativity. For example, answering an unconventional question, writing a poem or generating an image, or navigating in unfamiliar terrain. It is neural networks that exhibit abilities that are usually attributed to AI in general.
Modern neural networks like Chat GPT or Stable Diffusion are Deep Neural Networks or deep learning systems. Their main feature is that several layers or levels of evaluation are used between the input and output during training and data analysis. Such networks can handle more complex tasks and work in several directions. For example, they can work simultaneously with images and sound, or regular text and program code.
To put it briefly:
- AI. A direction of computer technologies that imitate human activity. Such systems are not always able to think and can work according to an algorithm, like a regular computer program.
- Machine learning. One of the directions of AI technologies. These are systems that learn on their own using a data set.
- Neural networks.Computer systems based on machine learning. They can apply the acquired knowledge to adapt to changing conditions on the fly and demonstrate creativity.
- DNN — Deep neural network. Neural networks can be useful in more than one direction and analyze information based on several features. For example, GPT can write ordinary text and program code, and perform calculations.
Overview of popular neural networks. Where can they be used?
Neural networks are everywhere now. They control self-driving cars from Tesla and Google, improve TV images, optimize photos taken with smartphone cameras, and monitor spam emails. The editors of Skillbox.by have prepared an overview of tools that are available to the average person and that can be used in work and the creative sphere without large capital investments.
Chat GPT
The most talked about neural network. This is a pre-trained text model that can answer questions and create content. Its capabilities don't stop there. Chat GPT can rewrite texts, translate articles into hundreds of languages, create business plans and recipes based on a grocery list, and write program code. The only question is quality. For example, the neural network is not yet capable of creating high-quality texts—Chat GPT creates stylistically complex texts with a lot of fluff; information must be double-checked, and the uniqueness of the texts must be manually increased.
The basic version of GPT-3.5 is available for free, but there is also a premium subscription with advanced features. Simply register on the website. Dozens of services like Jasper.AI or hypotenuse.ai also operate on the basis of GPT.
Bing AI
It is also Microsoft Copilot. A search chatbot based on Chat GPT 4.0. In addition to the Chat GPT functions, it can find data on the Internet, make a short summary of the information published on the page. The service is free, available in the Bing search engine (in the Edge browser), the Bing mobile app, and the Windows search bar. It is also used in Skype and other Microsoft applications.
Mid Journey
A generative neural network trained on millions of images. It can generate images based on a text description in English. Supports various art styles. It can create an image in the style of anime, photorealism, futurism, cyberpunk, and even paint a picture imitating the work of a specific artist. The service is available through Discord with a paid subscription.
Stable Diffusion
Although the neural network is inferior in popularity to Mid Journey, it is actively used by marketing companies, production companies and even film studios. Unlike previous neural networks, Stable Diffusion is available as open source. It can create not only images, but also videos, music, sounds, and sound effects. You can download it yourself, assemble it and train it on your PC.
There are also online services that work on the basis of a neural network: clipdrop.co, the official studio from the developer dreamstudio.ai, stablediffusionweb.comThere are also independent projects that were developed based on the source code. For example, unstability.ai.
Like GPT, the development of Openai. A neural network generates images based on a text description. Available for free with virtually no restrictions in the Bing search engine in the "Images" section.
Free image generator. Generates portraits of people for avatars, images of animals, vehicles, and landscapes. Many marketers use the service as a free photo stock to fill advertising profiles or create banners.
Among the interesting projects, we can also note:
- SlidesAI is a plugin for Google Slides that can quickly create a presentation based on added material.
- Yandex SpeechKitis a free cloud-based application that can recognize speech, translate it into text, and vice versa. You can customize the voices.
- Synthesia is a universal service that can generate videos based on a text script and voice the text. More than 120 languages are available, a huge selection of voices.
- Runway Gen‑2 — a video generator based on text description. The service is based on the Stable Diffusion neural network. It allows you to create commercials, educational videos, and cartoons. Available with a paid subscription, but there is a trial version - you can create three four-second videos for free.
The Future of Neural Networks. Useful resources from the editors of Skillbox.by
Neural networks are developing rapidly. Existing generative models are getting better, new capabilities are emerging. Hundreds of new AI services are also appearing. For example, an AI editor that can cut out and replace the background, increase the image resolution, and add additional objects to a photo is already built into Photoshop, Pixlr, and Canva.
Neural networks are not yet perfect and cannot completely replace humans. AI can handle a simple banner, image, entertaining post for social media, or product description for a marketplace. Not without proofreading, of course. It is impossible to rely on AI when writing articles and longreads, biographies, and publications about history. Artificial intelligence often confuses dates or simply makes up information, even about famous people and events. More complex tasks, like generating videos, require the use of multiple services—for creating the script, infographics, and voiceover. Image generators are also imperfect. They still can't process text and often struggle to render hands correctly.
But this does not mean that neural networks are useless. They can be delegated routine tasks that will make work and life easier. Therefore, the editors of Skillbox.by recommend not to be afraid that neural networks will replace people, but to learn to use them where possible, increasing your competitiveness.
Literature:
- Y. Goodfellow, I. Bengio, A. Courville "Deep Learning";
- D. Rutkovskaya "Neural Networks, Genetic Algorithms, and Fuzzy Systems";
- Rashid Tariq "Creating a Neural Network";
- Yann LeCun "How Machines Learn: The Revolution in Neural Networks and Deep Learning."
Internet Resources:
- Neural Networks — a blog dedicated to generative neural networks for creating images.
- OpenAI — a portal for the company that created DALL-E and GPT. There's a lot of information and documents on neural networks here.
- Neuroset.Info — Russian-language forum about neural networks.
- Machine Learning Mastery is a blog dedicated to neural networks and machine learning.
The Skillbox.by editorial team also recommends reading:
Don't equate admission to the end of the world. Expert advice for teenagers and parents
