A Neural Network

A Neural Network

What is it about?

A neural network picks the optimal output from a list of choices. This has many practical implications. Traditionally it would be performed by a human but now there is machine learning techniques which do this very well. The network can work directly on raw data. For instance images which makes it fast with little manual intervention.

For instance a claims payments process include input from the policyholder and the neural network can decide whether the claim is immediately payable or needs a review. In one instance when this was tested 25% of the claims where paid within 3 seconds.

Another application is whether to accept a life insurance policy right away or send it through to an opinion from a medical underwriter. If the network can approve a large share of the application directly this will drive efficiency within the company.

In addition, neural networks are being used by insurers for prediction of claims. This has been used for both frequency and severity of claims and for development of loss triangles.

The tools have seen increased interest from actuaries, risk managers, underwriters, and other insurance professionals. National actuarial organisations are actively looking into this area as well.


  • Low barrier from development to business implication.
  • Applies to many common business process in use in insurance companies today.

Deep Learning

Neural networks is closely related to deep learning, one of many machine learning techniques. The “deep” in “deep learning” refers to the number of layers through which the data is transformed.

Large-scale automatic speech recognition is the first and most convincing successful case of deep learning. The initial success in speech recognition was based on small scale recognition tasks. The data set contains 600 speakers from the major dialects of American English, where each speaker reads 10 sentences.

3 steps to set up a neural network

  1. Frame the problem. Describe the problem in computer language
  2. Get the data. Decide what data is the right to use.
  3. Train the network. This can take various levels of complexity. Depending on the problem.
  4. Validate the results. Continue develop the process.


This area is developing fast so watch out for the latest news. Learn more for instance here, http://neuralnetworksanddeeplearning.com/


This Post Has One Comment

Leave a Reply

Close Menu