Artificial Intelligence: The Neural Network and How It Works

March 29, 2021 • Zachary Amos


Artificial intelligence (AI) isn’t just one kind of tool or technology. AI developers utilize a range of approaches to create software that solves problems closer to human thinking than conventional, rules-based programming.

One of the most powerful and resource-intensive AI strategies is the neural network. This is a kind of machine learning that uses a model loosely based on the way the human brain works. Developers can “train” a neural network to create a highly effective pattern-finding algorithm that’s excellent at tackling unstructured data.

Here’s how artificial intelligence neural networks work — plus how they may change data analysis and tech in the near future.

The Basics of Neural Networks

Neural networks are a kind of AI, meaning they’re intended to replicate human thinking to solve problems and answer questions.

What sets neural networks apart from other kinds of AI is that these networks are a type of machine learning. Unlike other types of AI, neural networks are designed to learn — adapting their statistical model based on new information.

Neural networks are loosely based on the human brain. In the network, developers pass data through several layers of interconnected nodes — or neurons. Starting with an input layer that contains training data or new information, the neural network passes information through a series of layers made of multiple nodes. Each node connects to every other node in the next layer. 

By turning a node on or off — or by adjusting the statistical “weight,” or importance, of each node — the network can create varied and complex output.

Eventually it arrives at the final, output layer and generates a finished product.

Training a Neural Network

Developers of an artificial intelligence neural network first train the network on a data set relevant to what they want to analyze. For example, researchers who wants to help scientists identify foliage may train their network on labelled plant images.

During the “teaching” process, the neural network will run through the input dataset, trying to build a statistical model that represents the relationships in the dataset. The neural network, effectively by trial and error, runs through the dataset.

As it does so, it attempts to provide an answer to a given question. For example, “what kind of plant is in this photo?”

During training a neural network may start to “see” patterns in the data. For example, it may find clusters of colored pixels that may mean variegation seen only in a certain species of plants. These patterns are fed back into the neural network’s algorithm through a process called backpropagation. As a result, the network is able to “learn,” and over time, start producing more accurate answers.

Once a network has finished training, you can give it new information. Then, it will use the algorithm it has developed to provide you an answer.

Why Researchers Use Neural Networks (And Why They Sometimes Don’t)

These networks are helpful because it effectively generates an algorithm that can identify characteristics in a set of data without being prompted first. If you don’t know what kinds of patterns may be in a data set, a neural network is extremely effective at helping you break down and organize that data to find patterns and relationships hidden within.

The approach, however, has its limits. 

Some experts call neural networks “black boxes.” Devs refer to the layers of nodes between input and output as “hidden layers.”

Even if you have access to the code they generate, you can’t really tell how a neural network is “thinking” — why it produces the outputs it does or what characteristics it’s looking for. If a neural network returns useful predictions, you may need to analyze these predictions to understand what kinds of patterns exist in the data.

This is changing — one Google researcher, for example, has started work on a tool that helps developers understand how the network “thinks.”

The networks are also fairly resource-intensive. Without a high-powered computer, you’ll spend a lot of time training your network and letting it run before you see usable output.

How Researchers Will Use Neural Networks in the Future

You’ve likely already encountered an artificial intelligence neural network in your daily life. Google uses the tech to power its search engine. Amazon does the same for the recommendations on its digital storefront.

In the near future, neural networks will be essential tech in self-driving cars. Neural networks are one of the best available methods for creating computer vision algorithms, or models that can analyze real-time footage. Computer vision can allow autonomous robots and vehicles to respond to video input on the fly to navigate safely.

Because neural networks can analyze any kind of data, so long as you have enough, they may soon have major impacts on a number of fields — from medicine to industry to commerce.