Neural Networks

Our neural network technology offers businesses the opportunity to take stock of detailed historic data and help make predictions for the future.

For some years now, researchers around the world have been investigating and perfecting Neural Networks for such diverse tasks as:

Financial and Insurance Risk Assessment and Validation

Some credit rating companies use neural networks to assess the risk rating of clients' customers.  Similarly, insurers may use the technology to determine policy pricing for different categories of client.

Industrial Process Control. 

Applications include furnace control and other processes where traditional mathematical models using second order differential equations come into play.

Embedded Applications

such as robotics, ABS braking systems, washing machines....

To summarise

Neural Networks excel in applications where varied external events can be modelled on historic data. It is unnecessary for the investigator to know anything about the solution of the model in order to achieve good results.  A good understanding of the business, process or task is however fundamental.

The way that this is accomplished is by selecting appropriate "events" in the historic data and using that to train the neural network through several iterative processes.  This is usually difficult and very time consuming.

Neural Networks are literally a simulation of a primitive brain - specialised exactly to your criteria.

The data does not have to match future events exactly, but given the right training data the network should converge on good results - often better than traditional mathematical models.  Neural networks may even achieve 100% accuracy within your criteria.

Once the Neural Network has been trained within agreed criteria it is "frozen" and can be embedded in a run-time application.  The run-time application is very efficient and may be compiled into  process control or embedded systems.

Our technology allows us to create run-time versions of trained neural networks compiled in the C and C++ programming languages to allow very fast execution and porting to embedded applications.

Our network training technology uses special unpublished training algorithms based on previous research.  We train the networks on our own custom test-bed using C++ and SQL server in conjunction with our own in-house class libraries including data caching enhancements.  This allows us to work efficiently with very large data sets and manipulate historic data, as we feel appropriate to solve the task.

The data does not have to match future events exactly, but given the right training set it is often possible to get excellent results very quickly.

It is important to note that not all predictive problems can be easily solved with this technology.  The consultative phase may take some time.  It is important in training to ensure that the network does not "memorise" every instance of the training data.  The final solution, if one exists is often a compromise that gives results within your criteria.  Bear in mind however that this may be much better than any existing mathematical model.  However, not all problems can be solved using these techniques.  

A feasibility study is required to give us indications of the likely success of the project.

“We’re here to develop solutions that are right for your needs today and tomorrow”.