New successes in a digital world require new approaches. Especially when processing large amounts of data, the limits of human imagination and the resources available for data evaluation, such as time, personnel, capacity and budget, are often quickly reached.
While simple statistical methods were used in conventional data processing, we at Organon Informationssysteme GmbH use so-called Artificial Neural Networks as the basis for our Artificial Intelligence processes. These networks are able to train the network through the application of various algorithms or their learning methods (supervised learning, unsupervised learning, reinforcement learning) and thus calculate your data faster, more precisely and automatically.
The Artificial Neural Networks (ANN) consist of an input layer, possibly further intermediate layers and an output layer. There are different structures of ANN. So-called single-layer feed-forward networks, multi-layer feed-forward networks and recurrent networks.
The single-layer networks consist of an input and an output layer and only forward the outputs in the processing direction (“feedforward”). Multi-layer networks, so-called multilayer perceptrons (MLP), consist of input and output layers, the so-called “hidden layers”, and can be both feedforward-controlled and recurrent. In contrast to the feedforward-controlled networks, recurrent networks also have backward-directed edges, so that the outputs of a step can again serve as inputs for neurons in the previous layers or within their own layer and thus increase the dynamics of the network.
There are different methods for learning in neural networks, for example supervised learning, reinforcement learning and unsupervised learning.
Supervised learning is based on training “labeled” data sets, i.e. data sets for which the correct results have already been assigned. Techniques of supervised learning are used in feedforward and MLP models. Learning in monitored models is also called “error backpropagation algorithm”. The network is trained in that the algorithm adjusts the synaptic weights of the neurons on the basis of the error signal, i.e. the difference between the calculated and desired output values. The synaptic weight is proportional to the product of the error signal and the input instance of the synaptic weight. Based on this principle, each learning epoch consists of two steps, a forward pass and a backward pass.
Reinforcement learning is a learning method that is applied to an agent in a dynamic environment that can take various actions to achieve its goal. Unlike other algorithms, a problem and a series of actions to be carried out are not defined. Instead, the network must determine those actions that promise the best results.
In order to make the success measurable for the agent, he receives a reward for executed actions. The reward occurs on the one hand when changing to a different state, on the other hand a total expected profit is calculated. The agent then tries to get a high reward through every action and to maximize the overall profit. The advantages of reinforcement learning in dynamic systems make it particularly interesting for use in games. A combination of supervised and reinforcement learning is the basis for the well-known AlphaGo, an agent who masters the board game Go and even defeated top international players.
The success of machine learning systems often requires very large data sets that have already been labeled and that are very difficult to obtain. Unsupervised learning is a learning method for inputs that contain patterns which are not or mostly not classified, which makes the learning process very difficult.
It is possible to access completely unevaluated data sets, for example large amounts of data from the Internet. In unsupervised learning, for example, structures such as corners and edges or structures from certain object classes (e.g. car tires or parts of a face) can be helpful for object recognition with the aid of a small, labeled data set.