What Is a Neural Network?

A neural network is a collection of neurons that take input and, in conjunction with information from other nodes, develop output without programmed rules. Essentially, they solve problems through trial and error.

Neural networks are based on human and animal brains. While neural networks are advanced enough to beat human opponents at games like chess and Go, they lack the cognitive abilities of a human toddler and most animals. The importance of this foundational technology was recognized when the 2024 Nobel Prize in Physics was awarded to John Hopfield and Geoffrey Hinton “for foundational discoveries and inventions that enable machine learning with artificial neural networks.”

Neural Network Elements

Neural Network Elements

A neural network is made up of densely connected processing nodes, similar to neurons in the brain. Each node may be connected to different nodes in multiple layers above and below it. These nodes move data through the network in a feed-forward fashion, meaning the data moves in only one direction. The node “fires” like a neuron when it passes information to the next node.

A simple neural network has an input layer, an output layer, and one hidden layer between them. A network with more than three layers, including the input and output, is known as a deep learning network. In a deep learning network, each layer of nodes trains on data from the previous layer. The more layers, the greater the ability to recognize more complex information, based on data from the previous layers.

The network makes decisions by assigning each connected node to a number known as a “weight.” The weight represents the value of information assigned to an individual node (i.e., how helpful it is in correctly classifying information). When a node receives information from other nodes, it calculates the total weight or value of the information. If the number exceeds a certain threshold, the information is passed on to the next layer. If the weight is below the threshold, the information is not passed on.

In a newly formed neural network, all weights and thresholds are set to random numbers. As training data is fed into the input layer, the weights and thresholds are refined to consistently produce correct outputs.

How Does a Neural Network Work?

Whether it’s biological or artificial, the power of a neural network stems from the way simple neurons are linked to form a complex system greater than the sum of its parts.

Each neuron can make simple decisions based on mathematical calculations. Together, many neurons can analyze complex problems and provide accurate answers. A shallow network consists of an input layer, a hidden layer, and an output layer. A deep neural network has more than one hidden layer, which increases the complexity of the problems it can analyze.

A neural network learns to complete a task by examining labeled training examples. The samples must be labeled so the network can learn to distinguish items based on visual patterns correlated with the labels.

A neural network has three functions:

• Scoring input
• Calculating loss
• Updating the model, which begins the process over again

A neural network is a corrective feedback loop, giving more weight to data that supports correct guesses and less weight to data that leads to mistakes. A feature known as backpropagation trains the network to identify correct responses and ignore incorrect responses.

Types of Neural Networks

Not all neural networks share the same architecture. The structure of a network is chosen based on the kind of data it needs to process and the task it needs to perform. Three types are especially important for data science students to understand.

Convolutional neural networks (CNNs) are designed for image and video data. By scanning input in overlapping grids, CNNs have become the standard tool for image recognition, medical imaging, and computer vision.

Recurrent neural networks (RNNs) are built for sequential data such as text, speech, and time series. Unlike feedforward networks, RNNs carry a form of memory from one input to the next, making them useful for tasks where order and context matter.

Transformer networks are the architecture now powering most large-scale AI systems — from language translation and search to AI writing assistants and code generation tools. First described in the 2017 paper "Attention Is All You Need," transformers process all parts of an input simultaneously using a mechanism called self-attention, rather than sequentially. This enabled training on vastly larger datasets and produced the large language models (LLMs) that underpin modern generative AI.

A Few Concrete Examples

Neural networks are primarily used to classify and cluster raw, unlabeled, real-world data. They work behind the scenes of familiar technologies, such as online image comparison tools and financial decision-making tools for large corporations. A neural network can also analyze web browsing histories to develop recommendations for users.

Classification

Neural networks typically excel at classification tasks, which require labeled datasets for supervised learning.

For example, neural networks can find visual patterns in thousands of photos and consistently apply labels at high speed. Through training, they become good at solving complex, confusing problems. The data scientist doesn’t have to program the neural network with characteristics to distinguish between dogs and cats; the neural network learns to distinguish the most important features itself.

A neural network can learn to classify data with labels that correlate with the information it can analyze.

A major recent example is Google DeepMind’s AlphaFold, a neural network that predicts the three-dimensional structure of proteins from amino acid sequences. AlphaFold’s accuracy in this task, which previously required years of laboratory work, was recognized with the 2024 Nobel Prize in Chemistry. The system has generated predicted structures for more than 214 million proteins, accelerating drug discovery and structural biology research worldwide.

Clustering

While they excel at identifying differences, neural networks also work well for clustering or detecting similarities. A learning neural network can analyze millions of data points and cluster them according to similarities. This can be applied to images, emails, voice messages, or news articles.

This capability is likewise useful for identifying anomalies, or things that don’t correspond with group characteristics. For example, clustering is used to identify unusual behavior—such as fraud—by identifying data that doesn’t correspond with the most common actions.

Predictive Analytics: Regressions

Classification and clustering produce static predictions, such as an image labeled "dog". That identification won’t change over time. Regression analysis enables neural networks to predict future states based on past events. A future event becomes just another data point.

For example, the neural network can read a sequence of numbers and predict the next most likely number. It can apply the same analysis to more complex events, such as predicting when a customer may leave a store or when a piece of manufacturing equipment is likely to fail.

Regression analysis forms the basis for predictive analytics. By using regression analysis, a data scientist can model the relationship between a dependent variable (the outcome) and one or more independent variables (the input). Regression analysis will reveal any significant relationships between the independent variables and the dependent variable, as well as the magnitude of that impact. In other words, when the independent variables change, how much and in what way will the dependent variable change?

A basic neural network uses linear regression to manage one input and one output. Multiple linear regression comes into play when there are many input variables. In this case, each node in the network performs multiple linear regression, weighting each data point as it passes through the layers. The net tests the inputs as it tries to reduce errors.

Each node acts as a switch, allowing or blocking the input from the nodes around it through the network. Non-linear regression passes the input through the network until it reaches the final layer.

Neural networks use techniques such as gradient descent and backpropagation to refine their algorithms and find the optimal model for the regression.

Application of Neural Networks

Neural networks are integral to the development of machine learning and artificial intelligence applications. At this point, they don’t approach the cognitive abilities of a 4-year-old child. But they are being used in self-driving cars, facial recognition, language translations, artistic endeavors such as creating new colors, and large language models that power generative AI tools.

The growth of artificial intelligence has been fueled by the lower cost of cloud computing and graphics processing units, which enable the processing of large volumes of images for training. The widespread availability of electronic images and other data already tagged with information makes training easier and faster.

The capabilities to classify, cluster, and make predictive decisions have boosted the integration of neural networks in scientific research, advertising, e-commerce, customer service, preventive maintenance, and many other disciplines. Neural networks scan images of the night sky looking for new astronomical details. Messaging filters intelligently separate useful and unwanted emails and voice messages. When linked with sensors, a predictive analytics system can forecast when a hydraulic pump on a manufacturing machine will need servicing before it fails.

Explore a Career in Data Science

As neural networks play a growing role in research and commercial applications, the need for data scientists to create and operate them will also grow. If you’re interested in data science, check out the career and degree program information available through the Master’s in Data Science.