Neural Networks and Deep Learning

Neural networks and deep learning

Deep learning technology is a promising field of knowledge that strongly influences the development of artificial intelligence, defining the boundaries of self-learning based on statistical databases, integration of functions, formulas.

This will increase the efficiency of decision-making, especially in critical industries. For example, in medicine, to accelerate the search, synthesis of new drugs, and improve diagnostic accuracy by classifying signs of identical diseases.

How can neural networks and deep learning help in this? What is needed for this, what methods are used, let’s figure it out.

What is it

The concept of deep learning of artificial multilayer neural networks is a set of methods and tools that include nonlinear transformations, multi-level abstractions for adding new functions, increasing processing speed, and accuracy of responses.

The concept itself has been generally accepted in the scientific community since 1986, but the work of Soviet scientists on a working solution for multilayer perceptrons was published in 1965. In the early 80s, Fukushima Kunihiko introduced the neocognitron, an object recognition solution.

The real research began in 2000, when there were sufficient digital capacities and volumes of datasets (training materials). And it entered the modern stage in 2010, thanks to new high-performance graphics processors and convolutional neural networks. Two years later, ImageNet competitions were held, including the following stages:

• Detecting real scenes in a photo;
• classification and localization of objects from 1000 categories.

The winning team was able to reduce the number of errors by 16% (compared to the average of the test samples), to 65%. Now the AI performs the same tasks with an accuracy of 94% and a speed far exceeding the capabilities of the human eye.

How neural networks work

Multilayer models consist of rows of grouped neurons connected to the previous and next layers. There are no connections between neurons within the layers themselves.

The layered structure looks like this:

• The first distribution or input layer does not process, but only accepts the elements of the vector (links) and passes them to the next layer for analysis.;
• intermediate (more than one), hidden – it resembles a black box, since calculations, assignment of weights according to the significance level of the feature that occur in it, the inputs and outputs are inaccessible to the user.;
• The output is a result (forecast, essay, art, or translation from a foreign language).

The more levels there are, the more accurately the program will make a forecast and recognize the object. At the first hidden levels, it will notice lines, dots, circles, on the next - paws, ears, their shape, color, length of hair, and at the exit - it will determine the breed of the animal.

This principle resembles the work of the human brain, albeit greatly simplified. Therefore, such models can be called artificial intelligence. For example, convolutional neurons were created based on knowledge about the structure of the human brain. Because they roughly recognize visual information in the same way. Considering not each pixel individually, but in conjunction with neighboring ones, which allows you to select individual fragments and objects as a whole. After analyzing and comparing the features, it gives an accurate answer - is it a car, a tree, or a house.

By the way, most of the processes in the hidden layers still remain a mystery - it is not known for certain how AI solves problems. Therefore, the idea of explicable AI, the development of which would be transparent and accessible, is increasingly being heard in the scientific community. This will affect the awareness of responsibility by developers, will contribute to legal regulation, the creation of standards regarding ethical standards, copyright in the field of technology.

To the question of whether AI can think. No, there are blind spots in the structure that are inaccessible to humans, but the machine does not have consciousness, cause-and-effect relationships, or real experience. Even if it passes Turing tests. Therefore, AI will not reveal a new chemical element, a physical theory. And all abilities come down to complex mathematical functions, formulas, vectors, and numerical values.

How to train

A person needs experience to learn something. For example, to read, you need to learn the alphabet. Programs can’t do this on their own, they need a human.

The approach and methodology depend on the complexity of the architecture. In machine AI, a teacher is needed who would “show”, for example, a tangerine and a banana, “tell” that a tangerine is orange and a banana is yellow. I have fed more than one hundred examples, manually checked the correctness of the answers. I repeated this sequence several times before the machine could find these fruits in the still life. If you add pineapple to the selection, you will have to repeat everything again.

Complex models do not need to explain anything, it is enough to provide structured or labeled data. During processing, she will find the signs that unite the objects, distribute the weights according to the significance of the representations and classify the objects. If you add, say, an orange or an apple to the selection, it identifies them too.

Data is adjusted to meet the requirements of the program, and Data Science specialists or ML engineers are involved in the configuration. Their tasks include:

• information collection – open datasets, data sets for identification of handwritten characters, classification of objects;
• architecture definition – depending on the conditions, it can be convolutional, recurrent;
• customization for specific requests – activation functions (for memorizing complex dependencies), losses, optimizer, time, batch size (the number of examples on which the AI will learn);
• testing, pre–tuning - error correction, increasing the volume of introductory if the introductory is not enough.

There are no recommendations on settings - the programmer determines the parameters himself depending on the complexity of the task. When using reinforcement methods, the AI receives encouragement when completing the task correctly.

Now about where Deep learning is used:

• marketing – consultant chatbots, text generation, development of personalized recommendations based on customer preference analysis, demand forecasting, evaluation of indicators, advantages, disadvantages of competitors’ product;
• security – cameras on roads, video surveillance systems at airports, train stations;
• medicine – auto-completion of documents, classification of symptoms based on X-rays, prediction of reactions to medications;
• self–driving cars - interaction with road users, identification of obstacles, understanding the meaning of road markings, route building;
• banking and finance – forecasting stock prices, return on investment, fraud protection, reporting, budget optimization;
• Creativity is the creation of paintings, comics, art, logos, prints, commercials, and creatives.

To understand how neural networks work, learn more about algorithms, and learn how to write code for them, you can read Michael Nielsen’s book Neural Networks and Deep Learning. This is a practical guide with examples, puzzles, and detailed instructions for beginners.

And access to neural networks is already available on the website. Without downloading a VPN, linking a foreign phone number, and with free test requests:

• Claude AI, an analog, competitor to ChatGPT, can generate document templates, write an essay or a story on request, recognize text, and can maintain a dialogue on any topic.;
• Gemini – recognizes and processes audio, video, handwritten numeric, alphanumeric characters, writes program code, suggests how best to answer questions in the messenger, can “support a conversation”;
• Flux, developed by Stable Diffusion, generates detailed photorealistic art based on a text description, combines photos to generate a new image;
• Ideogram – works with images, accurately integrates text into a picture, and is suitable for SMM specialists and marketers to create content;
• Midjourney will help you create 3D illustrations for animation, games based on the description (you can choose from several presets in the style of realism, anime, and others), posters or logos with the correct display of numeric and alphabetic characters, and the built–in algorithm will improve the quality of prompts.

To get access, you need to register on the website by following the prompts. You can use the free versions with a limited number of requests or subscribe by choosing one of the tariffs.

Tools

Tools are used to expand the possibilities of self-learning:

Python is a programming language in which programmers set parameters.;

Jupyter Notebook – an application for setting up;

TensorBoard is an application that visualizes metrics, structures, and simplifies the understanding of the principle of AI architecture;

Ray Tune is a framework that automates the selection of hyperparameters that AI uses as a basis;

As well as Microsoft Azure Machine Learning and similar services that provide cloud storage. That is, it frees you from installing software on a PC.

In addition, neural networks are used:

CNNs are convolutional systems that learn to identify patterns and key parameters (stripes, the shape of a tiger’s face) using mathematical operations of the same name. They are used for image processing.

RNN – can store information in memory, the same previous words in a sentence, as a person does while reading. Understands the context and emotional coloring of the text. It can be used for audio generation, machine translation, and natural language interpretation in areas where sequential processing is important.;

GANS are hybrid and consist of two parts. The name very accurately describes the principle of operation. The first part is that the generator creates, say, an image. The second is a discriminator that compares the generated image with the real one. As a result, it really turns out to be a competition. For example, the generator tries to bring the conditional image of a sea sunset as close as possible to photographic accuracy, and the discriminator tries to find inconsistencies.

Depending on the type of program, one tool or algorithm is not enough. AI like Ideogram may require convolution, hybrid algorithms, and decoder programs to convert visual information into vector information.

Results

Deep learning allows you to significantly expand the scope of application and understand how AI works. In addition to advantages such as performing routine processes (the same call center robots, advanced recommendation systems), AI will be able to speed up the drug testing process, simulate conditions, and predict the results of experiments in the field of nuclear energy. To help scientists simulate chemical interactions that cannot be replicated in reality.