In an advanced state, the neural network, through training and the analysis of recorded data volumes linked to a result or action —for example the record of purchases made with a customer's credit card — could consider the weight of certain features and variables to predict the most likely result. This allows the neural network to detect predictive patterns in new data as more real-time data is entered.
Recent breakthroughs in Big Data analysis and artificial intelligence have improved our ability to build neural networks that can start being created with raw and unlabelled data and obtain a considerable volume of feature detection unaided. This type of neural network is often known as "deep learning" as it requires lots of hidden layers to gradually extract these features. So far, most of the research and development surrounding "deep learning" has been focusing on the analysis of raw data, such as video, voice and text. This is why these techniques have aroused the interest of big players such as Google and Facebook who have large amounts of raw and unlabelled data through which they could gain greater business return or improve the service if they were able to classify them automatically.
In the second case, self-learning analytics techniques are models that train themselves as they go to detect and recognise relevant data patterns in the real-time data received. This analysis model includes several techniques based on artificial intelligence and finds inspiration in the way the human brain is constantly processing and analysing the information perceived in our environment to define our notion of the surrounding world.
Among these techniques are those based on adaptive models that allow us to fix an additional analytics layer to a neural network and detect new fraud techniques much earlier; the atypical self-calibration multilayer model used by FICO for its cybersecurity solutions that allows us to detect suspicious behaviours in order, control and execution relations, for example, to detect bots and other malware infections; and self-organising genetic algorithm models that are based on learning and reinforcement methods inspired by, not so much how the brain works as how the collective intelligence of a species adjusts to a changing environment through experimentation and natural selection.
Lastly, neuro-dynamic programming is an analytics methodology that allows us to anticipate the way present and future actions could contribute to a reward accumulated in the long run. This technique, also based on artificial intelligence reinforcement learning methodologies, simulates the way the brain learns complex task sequences through pleasant or unpleasant feedback that could happen later on.
In the company-customer relationship, the result of this relationship will depend not only on the latest decision made, but also on the series of decisions made along the relationship sequence. For industries such as the insurance industry, neuro-dynamic programming could change the way underwriters will work in the future. Insurance companies can use analytics to determine how lifestyle patterns will impact policies and premiums. Therefore, it is critical to make progress in analytics and artificial intelligence techniques that can think far beyond the current situation, as they would allow us to improve a business' operational areas considerably to see beyond the immediate consequence ahead of the next decision.
Sign up for CIO Asia eNewsletters.