Convolutional Neural Networks

Convolutional neural networks

The improvement and study of new approaches to the implementation of computer vision technology began when standard models could not cope with the classification of handwritten digits. It was a challenge. They tried to solve the problem by implementing FFNN direct distribution. But they were limited in terms of accuracy due to the inability to capture neighboring areas.

The situation changed in 1990, when Jan Lecun introduced his LeNet-5. It was the first bundle. Let’s try to figure out how algorithms work, how they are trained, what they are used for, and where they are used.

What is it

CNN or convolutional neural network is a logical development of the ideas of the perceptron, developments at the junction of biology and computer science, which shows partial (and the best among algorithms for recognizing visual information) resistance to distortion. For example, changing the angle when searching for an object in a scene, changing the scale, or shifting.

They are considered a breakthrough in the research of SNN technology. The first big statement when the model was noticed was Alex Krizhevsky’s victory at the ImageNet competition in 2012. Received for updating the record of reducing the classification of errors to 15%. At the time, when the participants’ designs showed an average result of 65% correct answers, this was indeed an achievement.

Features – architecture that combines three ideas at once, the ability to understand the local context. Provided that the nodes being processed are located in the space in close proximity, they contain homogeneous data. That is, having a grid structure - visual images.

The SNN architecture consists of layers:

• convolutional - here artificial intelligence removes unnecessary elements, leaving details that will help classify the object, getting a feature map at the output (you can set parameters by texture, shape or color, it will select the necessary ones itself);
• pooling - reduces the size of maps by removing minor features and simultaneously reducing the load of computing power; if necessary, after pooling, you can skip through the first layer, repeating the algorithm of actions;
• batch normalization - to put it more simply, converting calculations to obtain an acceptable average;
• fully connected - with the simplest communication system that provides the result of the analysis.

For example, in a photo, in any picture, each pixel contains information about brightness and color. By comparing neighboring connections, the AI will see the passenger car, on each individual element, and on the overall picture. Only, in order for the neural network to give the correct answer, you need to run the drawing through the layers several times.

First, she will remove unnecessary lines, then classify identifying representations - headlights, wheels. After that, he will proceed to more complex ones (radiator grille, roof, doors, which are different for trucks and cars), and the decision that a car, bird or house is drawn will be made in a fully connected row.

The principle of operation

It resembles the mechanisms occurring in the visual cortex of the human brain. For example, a person first sees a house, only then begins to consider the details - the color of the walls, the roof, the shape, the number of windows. After that, he evaluates, say, curtains and notices pots of flowers on the windowsill.

This was scientifically proven by the experiment of Hubel and Wiesel, who proved that some of the nerve brain cells were activated in parallel with the visual perception of vertical, horizontal, and diagonal orientation.

After research, scientists found out that these cells are assembled into a core structure, which is responsible for the connection - image-understanding-evaluation of the object that a person sees. It was almost a guide to action. The idea of simulating the actions of nerve cells was realized in the game changer LeNet-5. Consisting of 5 rows sequentially processing fragments of an object.

Modern programs now do the same thing - first covering the whole, then breaking it down into particulars. The architecture of the bundles, of course, differs from the biological prototype and vaguely resembles a funnel through which each element is gradually sifted.

If we talk about how convolutional neural networks work, then the process can be divided into 2 stages:

• preparation - artificial intelligence sees the image as a three-dimensional matrix, numeric array, and before uploading it, it is run through a special system to make the data readable for the program.;

• splitting into individual pixels, changing the center and neighbor values by the coefficient of the filter matrix – adding the resulting products and replacing the main representations.

After that, the order of mathematical operations is repeated for each pixel until all adjacent areas are processed and a new matrix with updated parameters is formed.

The accuracy of recognition depends on the decision and training. The number of incorrect results can be reduced by applying classical machine learning methodologies for perceptrons:

• gradient descent;
• forward - the input data goes through several rows, the mathematical function redistributes the features according to their significance, making the source more abstract, which is better perceived by artificial intelligence.;
• loss function is a loss function that calculates the difference between the predicted and true values.;
• reverse propagation of the error.

The latter is useful for adjusting weights in hidden layers, since it allows you to change the value of a neuron in direct proportion to the error of neighboring ones.

Computer vision specialists teach models by adjusting the neural network to a specific type of task. But more often, pre-trained AI is used in various fields, which work faster and more accurately. Plus, their performance is noticeably higher.

Advantages and disadvantages

CNN is good at image processing. They can recognize images and enhance photos. They are adaptive, integrated with recurrent AI for text processing, which expands the scope of applications of convolutions to:

• text generation;
• maintaining dialogues in the same voice or chatbots;
• processing of heterogeneous information - audio, video.

The disadvantage is narrow focus. They can’t read content if the elements of the whole are far apart. Therefore, without additional tools, they will not be able to read or create a summary of the text. It’s the same story with tables, because they either contain dissimilar values - numbers, letters, symbols, or the distance between the components does not match.

In addition, AI is demanding on the size of documents. If you offer the same drawing on A3 and A4 sheets, the SNN may not understand that they are identical in content.

Which neural networks can be used?

AI for recognition and generation usually consists of two algorithms. Convolution and result converter. Created on the basis of CNN:

1. Midjourney generates content, can be used in graphic design and drawing a series of illustrations in a single style, and will be useful for users who are interested in drawing and designers.
2. Ideogram - understands descriptions, allows you to choose a drawing style from 6 built-in presets, automatically optimizes prompt due to the integrated function and is able to insert readable numeric characters without blurring the background.
3. Gemini is a multimodal AI that can invent stories, draw from descriptions, maintain dialogue and process videos, and competes with ChatGPT in terms of information processing speed and functionality.
4. Flux - generative AI will create a drawing in detail comparable to a high-resolution photo, take into account the conditions of the print, and combine several photos into one.

Access to these and other AI is already available on the website. Without downloading a VPN or registering a foreign phone number. To start using all the features of AI, you just need to register and choose a free or one of the paid tariffs.

How they are applied

Convolutional Neural Networks, with a grid topology, are effective for identifying and classifying objects. Due to the ability to automatically “get” significant parameters, they can be used:

1. It is used as a diagnostic tool in medicine in the fields of radiology, cardiology, and dermatology. SNN analyzes the retinal image to identify early signs of changes associated with age-related degeneration and diabetes. Patterns on X-rays and MRI scans characteristic of certain organ lesions are revealed. They help to identify new symptoms and diagnose life-critical conditions at an early stage.
2. For working with documents - classifying passport scans, for example. Provided that the paper is loaded in the same size.
3. Security systems include cameras on the streets, at airports, at train stations, and devices for authenticating identity in a bank.
4. Generating creatives, visualizing design projects - AI can come up with an idea, improve an outline, add details to a product card for marketplaces and generate an image of the product itself.
5. In the transformation and interpretation of photo and video files from surveillance cameras, including unmanned vehicles. Yandex uses such AI in its self-driving taxis. The convolution processes the video stream from the camera, monitors the situation on the road, sees the markings and detects traffic lights, ensuring safer movement in the stream. The same technology is included in the equipment of unmanned aerial vehicles.

In the work of the SNA, they found application in generative art. When the AI creates an image, it edits a video or photo based on the text description. You can recall a project when Midjourney showed what Russian cities would look like if they were human. These images flew all over social networks, flashed on news portals and social media search feeds.

Some programs can copy the characteristic style of famous artists, generate anime-style art, and add text to an image. They can also be entrusted with the development of prints and logos. For these skills, they are appreciated by marketers and SMM specialists who use convolutions as a working tool for creating content.

Results

CNNs are improving, developing, becoming faster, more productive. They are used to create functional systems. This explains the success of ChatGPT, which can perform several operations, unlike generative models - communicate consciously, answer questions, and even joke. Write code on prompts or suggest how to fix a mistake, draw and create document templates.

In the future, with the advent of new, more powerful, high-performance structures, it is possible to expand applications for unstructured data types, fields, and industries.

The technology has already been introduced by global giants. Google is used to search for user photos, and Amazon is used to generate banners for product recommendations. And Pinterest is for the possibility of personalization of the main page.