What is a deep neural network?
Valérie Bécaert Valérie Bécaert
November 15 5 min

What is a deep neural network?

Everyone who has taken high-school math knows that humans are far from perfect computers. Yet while math may be hard, we excel at the kind of pattern recognition that eluded computers for many decades. A child can recognize dogs as dogs, despite the many different breeds, while computers are still asking us to update when we’re halfway through a movie on Netflix.

Yet the human brain provided the inspiration for the algorithms powering today’s advances in artificial intelligence: neural networks. Complex neural networks with many layers of processing are also known as deep neural networks, and they power the deep learning algorithms behind some groundbreaking AI work.

Brain inspiration

Our latest understanding is that the human brain — our own built-in computer — works by transmitting electrical impulses between a series of specialized cells called neurons, which are connected by pathways known as synapses. These electrical impulses are the format in which our brain encodes information. When information is passed through the brain, different selections of neurons are “activated” to handle it, which determines almost every aspect of our thoughts, actions, and behaviour.

Some of the most-used AI algorithms are inspired by this interplay of neurons and information, our own personal “neural network.” So it’s only fitting that they are referred to as artificial neural networks. These AI models have an approximation of neurons, individual calculation units, which work together to form layers between which information is passed in the process of analysis and decision-making. These artificial neural networks powered the initial gains in AI from earlier this decade that kicked off the modern AI boom.

Layers of meaning

It’s tempting to take the neural network metaphor too far — even though they are inspired by the human brain, AI models do not think or process data like we do. The human brain is massively complex, many times more complex than even the most sprawling neural network in use today, and humans have specialized skills developed over millennia of evolution. AI models are algorithms, mathematical functions. To say they “think” is incorrect — but they do learn.

The neural network architecture has some implications for how AI models process data. The first stage of the process is known as the input layer (in which we provide the data to the AI), followed by the hidden layer (where the neural net does its analysis) and the output layer (where it provides the results of its calculations).

An artificial neural network with only one or a few of these hidden layers performs what’s known as machine learning. But when there are many more hidden layers, it’s able to accomplish far more complex, deeper analysis – deep learning – and is therefore a deep neural network. Now let’s consider what deep learning means.

Deeper machine analysis

A deep neural network can make much more considered and accurate judgements than less advanced neural networks because it can apply so many more levels of analysis. There are many types of neural networks, including those that only go one way, with one layer passing information to the next, and those that use loops or other forms of communication between layers to have a kind of “memory” that allows them to make decisions over time.

Because a deep neural network has been trained with large quantities of pre-labelled data, when it receives new, unlabeled data, it knows what to look for in order to classify it. Deep learning depends on this combination: extensive knowledge and a deep neural network with which to handle it and process new information.

Real-world applications of deep neural networks

Deep neural networks are at the forefront of technologies such as face and voice recognition, and in any field that deals with analyzing and categorizing sensor data such as still images or video and audio streams. From speech-activated virtual assistants to autonomous vehicles, if a technology requires the AI to train using mountains of data from the physical world in order to become proficient, then it will likely use some form of deep neural network.

These sophisticated mathematical models are changing industries and sectors across the board. In the medical field, they have been developed to recognize signs of cancer, and other diseases, by analyzing photographs of patients’ blemishes. Having been trained on many labeled pictures of malignant melanomas, as well as images of benign moles, deep neural networks has become highly adept at telling the difference. This is the kind of work at which deep learning excels.

There are countless other examples. Insurance and finance AIs use their learned knowledge of fraudulent vs legitimate claims to assist fraud detection personnel. Likewise, manufacturers can now spot and even predict product quality issues. And, in the retail realm, deep learning can provide far more accurate forecasts of future demand than ever before. More and more applications are appearing all the time.

Exploring new areas

As well as helping us understand our world and facilitate advances and innovations in our lives, deep learning is even helping us learn more about the universe beyond.

With the help of deep learning, astrophysicists are discovering “gravitational lenses” that allow them to see further into space and understand our universe in greater depth than has been possible until now. AI has tirelessly taken on a task too big for human scientists and is covering new ground with speed and precision.