Is the Artificial Neural Network the Future of Tech?
Neural networks are among the central technologies of artificial intelligence. They’ve been around for years (you use one every time your smartphone uses facial recognition) and in dozens of applications (finance, business analytics, and enterprise training). But they’re growing ever more prevalent.
But what is a neural network, and how are they changing the face of tech? How are artificial neural networks changing the internet, and what do you need to know about them? These brain-simulating computer applications are easier to understand than you may think, and they’re delivering unprecedented processing power and new, exciting benefits.
What is a neural network?
A neural network is a machine learning tool that trains computers to “think” like human brains, such as the facial recognition in your smartphone camera. Also called artificial neural networks (ANNs) and simulated neural networks (SNNs), they learn by analyzing a series of training examples.
ANNs use artificial neurons (called perceptrons), programmed into their software. Each perceptron is a mathematical function that takes in numbers and outputs a result. These perceptrons interact in special ways to help the software learn.
Today, applications increasingly use ANNs. Some examples of neural networks:
- Voice-to-text
- Smart assistants like Google, Siri, and Alexa
- Online check deposit
- Online shopping search
- Facebook tagging and fake news detection
How do neural networks work?
The software “neurons” in a neural network form different processing “layers.” Each network bundles groups of these perceptrons into 3 types of layers: an input layer, hidden layers, and an output layer. Data comes in through the input layer, is processed in the hidden layers, and exits through the output layer.
The hidden layers of the ANN perform a simple but groundbreaking trick. Each hidden layer processes a different part of the data, and each “neuron” works independently on its given data chunk. Together, the neurons in a layer contribute to a single result.
How is a neural network like a human brain?
Our brains use layered processing, just like neural networks. When you watch a car drive down the road away from you, your brain’s “layers” must process:
- Edge detection
- Shapes (circles, squares, etc.)
- Color
- Size
- Motion
- Context
Your brain’s neural network layers process all those things independently, but each layer passes its results up the chain.
An artificial neural network’s layers work in the same way. Each layer owns a different part of the data, working on it in concert with the others to produce a unified result. Behind the scenes, different mathematical functions process the smaller chunks of data.
ANNs vs traditional programming
The way neural networks work is very different from traditional computer processing, where all the machine’s responses follow a set of pre-written rules (a program). Traditional programming might compare two images to see if they’re exactly the same. A programmer writes code that tells the computer how to check each pixel in each image.
By contrast, a neural network can tell you if two images are similar. It can learn to compare the images based on certain criteria, but it can also learn to decide on the criteria itself.
Neural networks vs machine learning
Neural networks go a step beyond even traditional machine learning. A machine learning model can “learn” from data, making decisions based on what it learns. But a neural network can do more, rearranging its own algorithms in the process.
So, while a machine learning algorithm may learn to recognize faces, an ANN may learn to get better at the process of recognizing faces, adjusting its own actions along the way.
What do neural networks do?
The applications of neural networks span multiple disciplines and industries, including enterprise planning, financial operations, product maintenance, and business analytics.
Where traditional computer programming is limited to inputs and outputs that the programmers expected, neural networks can work with unexpected inputs and still deliver accurate results.
Neural networks go beyond databases
For instance, traditional programming can recognize a printed number “7” or a letter “t” based on comparisons to image files stored in a database. But there are drawbacks to this process.
- The program must continually re-check the database, using large amounts of processing power.
- If the program encounters a font that is not in the database or a sample of human handwriting, it may not be able to process the correct result.
Neural networks process unexpected inputs
By contrast, a neural network can learn to recognize a number “7” or a letter “t” in any font or handwriting, based on learning from analyzing hundreds of examples. The more examples a network sees, the better its pattern recognition skills. The network won’t need to refer back to a database once it has learned the skill.
In the same way, a neural network can learn to “know” it’s looking at a face – even a face it’s never seen before. It can also learn to recognize financial trends, probability of machine failure, and other patterns traditionally hidden in a sea of data.
ANNs bring big data down to size
This ability to draw insight from large data sets becomes increasingly valuable as the flood of data grows exponentially. In 2021, our world will generate 79 zettabytes of data. That’s a 1 with 21 zeroes. That’s up drastically from just 2 zettabytes in 2010. By 2025, we’ll more than double today’s rate at 181 zettabytes per year.
Neural networks are increasingly skilled at navigating that data flood and drawing conclusions needed for economic forecasting, business planning, healthcare cost trends, and hundreds of other applications.
Training neural networks
Neural networks learn by using training examples. To train a network how to recognize a number “7” no matter the font or who writes it, developers show the network hundreds of 7s written in all different ways.
Each time the network tries to recognize the 7, it delivers a result. The result is evaluated by a trainer, and an error value is sent back across the network or backpropagated. The network uses the error value to adjust itself.
Neural network architecture
When a perceptron in an ANN processes data, it uses a value called a weight. With each processing attempt and error value, the perceptron adjusts its weight. As the network gets better at recognizing the 7, the adjustments to the weights get smaller and smaller, and the network’s skill improves.
By going through this incremental training process, the network stores the information that lets it draw insight from data in its own neurons and their weight values. That’s why a neural network doesn’t need to check a massive database of stored 7s written in different ways every time it wants to recognize a number 7.
The future of neural networks
Software developers continue to push back the boundaries of what neural networks can do. Think of the internet before Google. As a searcher, finding information was a needle-in-a-haystack chore. Google made the internet into an incredibly useful place where almost any question can be answered instantly.
In the same way, ANNs will continue to pervade every aspect of our lives, from doctor visits to banking, and entertainment to GPS navigation. If you notice things getting easier in the next few years, you may need to thank an ANN.
Types of neural networks
As neural networks find their way into a growing range of applications, they’re adapted into an exponentially increasing number of types. Here are just a few:
- Perceptron: The most basic, smallest unit of an ANN
- Feed Forward Neural Network: A one-directional network
- Radial Basis Network: Standard input-hidden-output neural network
- Deep Feed Forward: A standard network with multiple hidden layers
- Recurrent Neural Network: An ANN that uses time-series data
- Markov Chain: Uses probabilities to switch from one state to another
- Deep Belief Network: Stacked networks that together speed up the training process
- Graphical Deep Convolutional Neural Network: An ANN that specializes in manipulating 3D images
- Echo State Network: A sparsely-connected network designed for use in engine control, vibration analysis, seismology, and signal forecasting
- Neural Turing Machine: An ANN that blends neural networks with programmable computing
For more information on the different types of neural networks see this neural network chart.
Examples of neural networks
You already use ANNs in dozens of ways. They’re making your life easier and more fun in the background all the time.
Here are a few neural network examples:
- Social media: Facebook uses ANNs to recognize your friends and family in your photos. Instagram auto-suggests emojis and hashtags based on what you type.
- Online shopping: Ecommerce sites and platforms like Shopify are adding advanced searches, personalized shopping, and sales forecasting based on neural networks.
- Banking: Banks are exploring neural networks as a way to detect fraud and improve digital customer service.
- Mobile phones: Image recognition, plant identification, and even early skin cancer detection are finding their way into our phones through ANNs.
- Finance: Neural networks are forecasting stock market returns, evaluating creditworthiness, and detecting financial distress.
- Manufacturing: ANNs can monitor product quality and model manufacturing processes.
- Machine maintenance: ANNs are moving into the manufacturing world, predicting when machines need maintenance and diagnosing issues.
- Business Analytics: Predicting consumer demand, optimizing inventory levels, and estimating freight arrival times are among the many neural network applications in business today.
Neural networks summary
Neural networks have already changed the world in drastic ways, but we’ve only scratched the surface. Today, we use them in facial recognition, Facebook tagging, text-to-speech, online check deposit, and hundreds of other applications. In the future, neural networks will improve processing, internet speed, and sensor monitoring, making our lives easier and more fun.
About the Author: Tom Gerencer is a contributing writer for HP Tech Takes. Tom is an ASJA journalist, career expert at Zety.com, and a regular contributor to Boys' Life and Scouting magazines. His work is featured in Costco Connection, FastCompany, and many more.
Popular HP Products
