Content: An exposé which helps you build a sound business understanding on the subject of artificial intelligence. The introduction talks about the founding fathers of artificial intelligence and their main ideas. Then the question is asked, how we define intelligence. Shortly thereafter, we take a look at how artificial intelligence can help us and the question is asked what neural networks actually are. Finally, the structure of neural networks is analyzed in detail.
Introduction: Although the term artificial intelligence has only been heard by the general public in the last couple of years , the first reflections on artificial intelligence were already collected in 1956 at the Darthmouth Conference. Today this conference is called the birth of Artificial Intelligence as an academic subject, with its founders and thus the founding fathers John McCarthy, Marvin Minsky, Nathaniel Rochester, Claude Shannon. The aim of this conference was to describe the aspects of the learning process and the characteristics of intelligence so precisely that a machine could simulate these processes. Even back then, people toyed with the idea of using a machine language, undertaking various abstractions, automating computers, making detailed considerations about the scope of arithmetic operations in order to also integrate neural networks, and incorporating a certain degree of randomness in order to promote creativity and to generate associated problem-solving approaches, which humans have not yet been able to determine themselves, using the computer. And with the ultimate goal that the machine can self-improve.
In order to better understand the topic, one has to deal with the question: What is intelligence? Intelligence is the ability to solve problems and survive. Thus an intelligent person is able to receive know-how and to learn from it. She must have instincts and intuition, and be able to develop an understanding. It can perceive various influences such as sounds, tastes, images, etc. and process this information. So consciousness is part of intelligence. In combination with memory, which helps to learn situations and behavior patterns in order to react appropriately. Another characteristic is the ability to adapt, for example to be able to take what has been learned into account when making future decisions. There must also be a certain understanding of complexity, because complex problems can only be solved with deep problem-solving skills. The important criterion of creativity, in which new synapses are formed that have never been there before, is also relevant. And it must not be overlooked that the ability to interact with one's environment, such as using resources from the environment, is an aspect of intelligence. Another aspect of intelligence is the ability to see, plan and develop strategies for the future. And last but not least, the formation of a culture is a sign of intelligence: passing on knowledge over generations, cooperation in order to achieve common goals for the good of the group, to create new, unprecedented problems - which would never exist in non-intelligent beings, and the associated desire to solve these problems, all of this makes up intelligence.
The above introduction has shown that humans want to create artificial intelligence in order to solve complex problems faster and more efficiently and thus want to remedy manual processes - which helps to scale. To achieve this goal there must something exist, that has the ability to generalize, i.e. to be able to abstract to a special case. Nowadays, however, machines still require too much energy compared to humans. So it was already at the Darthmouth Conference that people thought about how this project would be approached, and it was decided that the system would be set up in the same way as with humans.
What are neural networks? Humans have neurons, which are electrically excitable cells that use electrical and chemical signals to record information, process it and then pass it on. The neuron is therefore one of the basic elements of the nervous system and is used to enable people to react to their environment: On the one hand, the stimulus is transported to the central nervous system by an ascending neuron, on the other hand, descending neurons, for example, stimulate the arm to move it away from danger. It is a simple structure for very complex tasks, because the typical neuron is divided into three parts: the cell body, the dendrites and the axon:
The cell body serves as the control center of the neuron. The heavily branched dendrites pick up the electrical impulses and pass them on to the cell body, so that it forwards the information through the axon to the next cell body.
Already at the Darthmouth conference, people thought about how to approach the artificial intelligence project and came to the decision that one of the structures can be taken over by humans: Neural networks could be used to create artificial intelligence. This knowledge could be adopted by Warren McCulloh and Walter Pitts, who already in 1943 thought about the connections of elementary units, which served as a kind of network similar to the networking of neurons and thus all arithmetic and logical functions could be calculated. Using this abstraction, a model for information processing is created that supports the assignment of connections to nodes. After such a neural network has been created, it is trained so that the network can learn.
How can a neural network learn?
On the one hand, learning can be promoted by developing new connections or by deleting existing connections. Another exciting aspect is changing the weighting (w) or adjusting the threshold values of the neurons, if they exist. The neural network can learn by adding or deleting neurons, or by modifying the activation, propagation, or output function.
Mainly, a network learns by modifying the weights (w).
How is the neural network structured?
1) We have x1 to xn inputs. (N is any number of inputs.) These are the input nodes. The inputs take the form of numbers and serve as activation values.
2) These inputs are weighted with w1 to wn. The weighting has hidden nodes because it is processed hiddenly.
All input nodes are linked to the hidden nodes:
The strength of the connections between the input and the hidden nodes is individually defined. It is similar to humans, in which those structures are more pronounced where there is more flow of information. If the value of the connection is high, then it means that the connection is strong. The activation values are passed on from node to node so that each node adds up the activation values it receives.
3) The resulting value from the additions is then modified by the defined transfer function.
What transferfunctions do we have?
- Uni step
- Piecewise Linear
4) The activation is then shown as an output. 5) At the beginning the question was looked at what intelligence is and what such a neural network is used for. In principle, it should help to solve problems, so the neural network is able to generate outputs based on inputs. It can therefore predict outputs and, among other things, it can also be that such an output is faulty. In this case, the difference between the predicted value and the actual value is passed backwards. The difference is divided into the weighting of each individual node, so that each node can take responsibility for this error according to its extent (e.g. gradient descent algorithm)
In the next blogpost we will dive deeper and analyze the topics of: Business Intelligence, Analytics, Advanced Analytics, Machine Learning and Deep Learning.