What is a neural network?

What is a neural network?

A neural network is a technique in artificial intelligence that teaches computers to process data in the same way as the human brain. It is a type of machine learning process called deep learning that uses interconnected nodes or neurons in a layered structure that resembles the human brain. He creates an adaptive system by which computers learn from their mistakes and constantly improve. In this way, artificial neural networks try to solve complex tasks, such as summarizing documents or recognizing faces, with higher accuracy.

What is the importance of neural networks?

Neural networks help computers make intelligent decisions with limited human input. They can study and model relationships between non-linear and complex inputs and outputs. For example, neural networks can perform the following tasks.

Summarize and draw conclusions

Neural networks can understand unstructured data and make general observations without special training. For example, they can recognize that two different input sentences have the same meaning:

• Can you tell me how to pay?
• How can I transfer money?

The neural network will understand that both sentences mean the same thing. She can also determine that Baxter Road is a place and Baxter Smith is a person’s name.

What are neural networks used for?

Neural networks are common in many industries. Among them:

• Diagnosis with Medical Image Classification
• Targeted Marketing with Social Media Filtering and Behavioral Data Analysis
• Financial forecasts by processing historical data of financial instruments
• Forecasting electrical load and energy demand
• Compliance and Quality Control
• Definition of chemical compounds

Below are the four most important tasks that neural networks help to solve.

Machine vision

Machine vision is the ability of computers to extract information and meaning from images and videos. With the help of neural networks, computers can distinguish and recognize images in the same way that humans do. Machine vision is applied in several areas, for example:

• Visual recognition in self-driving cars so they can react to road signs and other road users
• Content moderation to automatically remove insecure or inappropriate content from image and video archives
• Face recognition for identifying people and recognizing attributes such as open eyes, glasses, and facial hair
• Image tagging to identify brand logos, clothing, protective gear and other image details

Speech recognition

Neural networks can analyze human speech regardless of its speech patterns, pitch, tone, language, and accent. Virtual assistants such as Amazon Alexa and automatic transcription software use speech recognition to perform the following tasks:

• Assistance to call center operators and automatic call classification
• Convert clinical guidelines into real-time documentation
• Accurate subtitles for videos and meeting recordings for wider content coverage

Natural language processing

Natural language processing (NLP) is the ability to process natural, human-generated text. Neural networks help computers extract information and meaning from text data and documents. NLP has several uses, including:

• Automated virtual agents and chatbots
• Automatic organization and classification of recorded data
• Business intelligence of long documents: e.g. emails and forms
• Indexing key phrases that indicate mood: for example, positive and negative comments on social networks
• Summarizing documents and generating articles on a given topic

Recommendation Services

Neural networks can track user actions to develop personalized recommendations. They can also analyze all user actions and discover new products or services that are of interest to a particular consumer. For example, Philadelphia-based startup Curalate helps brands convert social media posts into sales. Brands use Curalate’s Intelligent Product Labeling (IPT) service to automate the collection and processing of social media user content. IPT uses neural networks to automatically find and recommend products that match the user’s social media activity. Consumers don’t have to rummage through online catalogs to find a specific product from a social media image. Instead, they can use Curalate’s automatic labeling,

How do neural networks work?

The architecture of neural networks follows the structure of the human brain. Human brain cells, called neurons, form a highly interconnected complex network and send electrical signals to each other to help people process information. Similarly, an artificial neural network is made up of artificial neurons that interact to solve problems. Artificial neurons are software modules called nodes, while artificial neural networks are programs or algorithms that computing systems use to perform mathematical calculations.

Basic neural network architecture

The basic neural network contains three layers of interconnected artificial neurons:

Input layer

Information from the outside world enters the artificial neural network from the input layer. The input nodes process the data, analyze or classify it, and pass it on to the next layer.

Hidden layer

Hidden layers receive input from the input layer or other hidden layers. Artificial neural networks can have a large number of hidden layers. Each hidden layer parses the output of the previous layer, processes it, and passes it on to the next layer.

Output layer

The output layer gives the final result of processing all the data by the artificial neural network. It may have one or more nodes. For example, when solving a binary classification problem (yes/no), the output layer will have one output node, which will give the result “1” or “0”. However, in the case of multiple classification, the output layer may consist of more than one output node.

Deep Neural Network Architecture

Deep neural networks or deep learning networks have several hidden layers with millions of artificial neurons connected to each other. The number, called the weight, indicates the connections of one node to others. The weight is positive if one node excites another, or negative if one node suppresses another. Nodes with higher weight values ​​have more influence on other nodes.
In theory, deep neural networks can map any type of input to any type of output. However, keep in mind that they require much more complex training than other machine learning methods. Such nodes need millions of examples of training data, rather than hundreds or thousands, as is the case with simple networks.

What types of neural networks are there?

Artificial neural networks can be classified according to how data is transferred from an input node to an output node. Below are a few examples.

Feedforward Neural Networks

Feedforward neural networks process data in one direction, from an input node to an output node. Every node in one layer is connected to every node in the next layer. Feedforward neural networks use a feedback process to improve predictions over time.

Backpropagation Algorithm

Artificial neural networks are constantly learning using corrective feedback loops to improve their predictive analytics. Simply put, it is about data flowing from an input node to an output node along many different paths in a neural network. The only correct path is the one that maps the input node to the correct output node. To find this path, the neural network uses a feedback loop that works like this:

1. Each node makes a guess about the next node on the path.
2. It checks if the guess is correct. Nodes assign higher weights to paths that lead to better guesses and lower weights to node paths that lead to incorrect guesses.
3. For the next data point, the nodes make a new prediction using higher weighted paths and then repeat step 1.

Convolutional Neural Networks

Hidden layers in convolutional neural networks perform certain mathematical functions (such as summation or filtering) called convolutions. They are very useful for classifying images because they can extract relevant features useful for recognition and classification. The new shape is easier to handle without losing features that are critical to a correct guess. Each hidden layer extracts and processes different characteristics of the image: borders, color, and depth.

How to train neural networks?

Neural network training is the process of training a neural network to perform a task. Neural networks are trained by first processing several large sets of labeled or unlabeled data. Based on these examples, networks can process unknown inputs more accurately.

Supervised learning

In supervised learning, data scientists provide artificial neural networks with labeled datasets that give the correct answer in advance. For example, a deep learning network learning face recognition processes hundreds of thousands of images of human faces, with different terms related to ethnicity, country, or emotion describing each image.

The neural network slowly accumulates knowledge from these datasets, which give the correct answer in advance. Once trained, the network begins to make assumptions about the ethnicity or emotions of a new human face image that it has never processed before.

What is deep learning in the context of neural networks?

Artificial intelligence is a field of computer science that explores methods for enabling machines to perform tasks that require human intelligence. Machine learning is an artificial intelligence technique that gives computers access to very large datasets for further learning. Machine learning software finds patterns in existing data and applies those patterns to new data to make intelligent decisions. Deep learning is a type of machine learning that uses deep learning networks to process data.

Machine learning and deep learning

Traditional machine learning methods require human input in order for the software to work properly. The data scientist manually defines a set of appropriate functions that the software should analyze. This limitation makes creating and managing software a tedious and labor intensive process.

On the other hand, in deep learning, the data scientist only provides the raw data to the software. A deep learning network extracts features on its own and learns more independently. It can parse unstructured datasets (such as text documents), prioritize data attributes, and solve more complex problems.

For example, when training software with machine learning algorithms to correctly identify an image of a pet, you would need to follow these steps:

• Find and manually tag thousands of images of pets: cats, dogs, horses, hamsters, parrots, etc.
• Tell the machine learning software which features to look for so that it can identify the image by elimination. For example, it can count the number of legs and then check the shape of the eyes, ears, tail, fur color, and so on.
• Manually evaluate and modify labeled datasets to improve software accuracy. For example, if you have too many images of black cats in your training set, the software will correctly identify a black cat, but not a white one.
• With deep learning, neural networks will process all images and automatically determine that they first need to analyze the number of legs and the shape of the muzzle, and only then look at the tails in order to correctly identify the animal in the image.

List of neural networks

List of more than 800 neural networks grouped by category. Neural networks for image, text, code, etc.

Architecture & Interior Design

Code

Gaming

Image

Music

Speech

Text

Video

Other