How a neural network learns to think for itself
The human brain is often compared to a computer. Yet in reality, the human brain is much more complex than any computer. But that has not stopped researchers from working to develop computers and software that act in ways similar to the human brain. One of these is a type of software called an artificial neural network (ANN). Neural networks are the systems that power artificial intelligence. So, what are artificial neural networks, and how do they function like the human brain?
An ANN is designed to simulate the way that cells in the brain communicate with each other. They are made up of interconnected processors, similar to how the brain is made up of billions of interconnected neurons. While a traditional computer uses a central processor to take instructions and do what it is told, an ANN uses many simple processors, called units.
The units in an ANN are connected in layers. First is an input layer, which receives data that the network will process. The data then moves through many other “hidden” layers. As it moves from one unit to another within the hidden layers, the data is weighted. The higher the “weight”, the greater the influence that each unit has on the next unit. The data with the highest weight has the most confidence, or accuracy. In this way, the network “learns” more about the data as it passes through the layers. Finally, the data passes to the units in the output layer. This layer then transmits the processed data to the outside world.
In order for ANNs to learn, they first need to be “trained” with a large amount of data. For example, if researchers want to teach an ANN how to identify an image of a dog, they must first input thousands of images of dogs. During the training, the ANN’s output is compared to the correct answer. If the ANN does not reach the correct answer about a particular image, it is instructed to go back through the layers and alter the weighting. This is known as back propagation, or deep learning, and is what makes an ANN network “intelligent”.
After the training, the ANN will reach the point where it can be presented with an image it has never seen before, and it will be able to determine whether or not the image contains a dog. In this way, the ANN starts to think like us. It recognises patterns and uses them to make decisions.
Neural networks can be applied to a wide variety of these pattern recognition tasks. For example, LinkedIn uses ANNs to detect spam. YouTube and Netflix use ANNs to make recommendations for what to watch next, based on viewing history.
There are some concerns about ANNs, however. They require a huge amount of data for training, and consume a great deal of energy. Also, because they can predict how people will act in certain circumstances, they can potentially be used to manipulate peoples’ behaviour.
But ANNs may soon be ubiquitous particularly as they become cheaper and faster. Already, they are useful in image and voice recognition, and are used in fields as disparate as security, translation, marketing, medicine and astronomy. At Springwise, we have covered the use of ANNs to create and evaluate art and to help e-commerce businesses personalise prices and promotions based on customer locations.
26th February 2019