How could Neural Networks help power the insurance industry?

The worlds of AI and machine learning are complex, vast and ever-evolving. So it’s no surprise that at Nuon AI, one of our guiding principles is a fascination for all things AI, not just the branch of AI our Price Adjust product uses – known as Reinforcement Learning.

Recently, the team has been spending time learning about Neural Networks, and exploring the potential this subset of machine learning could hold for future AI products to power the insurance lifecycle. 

Led by co-founder Mat Donfrancesco, and with thanks to Steve Roberts for sharing his insights, read on to find out what we discovered in our exploration of Neural Networks.

Further reading: How AI is transforming the insurance industry

What is a Neural Network?

An important starting point is to define exactly what Neural Networks are. However, that question is not easily answered. The field is very broad and depends heavily on mathematics and statistics.  Many resources online and in literature do not give a fundamental understanding of Neural Networks, but jump straight into the programming and mathematics behind them.  

It is easy to get caught up in the terminology: Linear regression, Stochastic Gradient Descent, Mean Square Error, distribution curves, integrations, activation functions using sigmoid algorithms…

Still with us?

To put it simply – Neural Networks, or artificial neural networks (ANN), are a subset of machine learning and at the core of deep learning. Their name and structure are loosely driven by the structure and processes within the brain, and the way that neurons communicate with each other. 

Structurally, they are composed of layers of nodes: an input layer, an output layer and usually zero or more “hidden” layers between the input and output layers.  Typically if the neural network has more than 3 layers, it is called a deep neural network.  There can be many layers of nodes, and each layer may have many nodes.

Each node in one layer of the network is connected to each node in the next layer of the network, except for the output layer which connects to the external application using the network.

A node can be thought of as synonymous with a neuron in a brain, connected to a number of other neurons.  It works by having mathematical “weights” and thresholds.  If the threshold is reached, the node sends an output to the nodes of the next layer of the network.  If not, then it does not send data to the next layer.  

how Are Neural Networks commonly used in tech?

You may not realise it, but Neural Networks are already a part of our everyday lives through the technologies that power many of our interactions, including…

• Voice Recognition tools such as Siri, Alexa
• Image recognition, including number plate identification or in social media for photo video manipulation such as highlighting a person’s face in a photo and applying texture to a background
• Car driving aids including auto parking, lane detection, and of course partial or fully autonomous vehicles. 
• Fraud detection tools used in the banking and insurance sectors 

Neural Networks rely on training data – and usually a large amount of it.  For example, suppose the network is to identify a dog in an image. In that case, it is “trained” by being “shown” many images (typically hundreds of thousands or even millions) containing at least one dog. 

Think how many types of dogs there are (breeds, colours, sizes), and then they can be in different positions, have different image backgrounds, be in colour, black and white, sketched or painted…the list goes on. 

The network is fine-tuned to identify dogs.  Images which have then not been used as input to the network would be used as test data.  This test data set would typically comprise images containing dogs, and images which do not.  The error rate of the network is determined to identify the accuracy and suitability of the network for identifying dogs from an image.