Neural networks in simple words

Neural networks in simple words

The world of artificial intelligence is advancing at great speed, as do all networks oriented to digitization. A good example of this is artificial neural networks (ANNs), a technology of great relevance for the future. But what is it and how does this great unknown works?

Artificial neural networks are mathematical models that seek to replicate the behavior of neurons in nature, organizing their structure in a way that emulates the brain. Considering that there are around 100,000 million neurons in this organ, it is an extremely ambitious and advanced technology.

 The objective of these networks is to achieve an intelligent system that can successfully perform complex tasks. Its operation is not oriented to be similar to that of a computer when it processes information but rather seeks to approximate artificial intelligence. Thus, artificial neural networks are capable of creating patterns, recognizing information, or solving complex enigmas, and can be very useful for sectors such as fashion or the fight against crime.

 In addition to this, neural networks have another important objective, which is to contribute to the creation of networks. As they are very fast in terms of information processing, they are capable of transmitting data at high speed between two points. As if this were not enough, with the passage of time, they acquire the ability to relate to each other, in such a way that they expand the available knowledge, renewing it and functioning autonomously.

 In short, it is a giant leap in the world of digitization, thanks to which artificial intelligence will have more and more prominence, taking charge of processes and making decisions in a similar way to humans. Although it sounds like science fiction, the reality is that advanced models are already being tested and implemented with great success.

How do Artificial Neural Networks work?

Despite their name, neural networks don't have a too complicated concept behind them. The name, as you can imagine, comes from the idea of imitating the functioning of the neural networks of living organisms: a set of neurons connected to each other and working together, without having a specific task for each one. With experience, neurons create and reinforce certain connections to "learn" something that remains fixed in the tissue.

 Neural networks allow you to find the combination of parameters that suits a certain problem in the best way. Neural networks are a model to find that combination of parameters and apply them at the same time. Speaking simpler words, the task is finding the best-fitting combination is "training" the neural network. An already trained network can then be used to make predictions or classifications, that is, to "apply" the combination.

 To fully understand how this network works, we are going to go with an example. Let’s suppose that you are students in a class in which the teacher has not exactly clarified how he is going to put the notes and what the criteria of estimation are going to be. To begin, let’s suppose that you have only two exams to take and you will have the grade for each of them and the final exam.

 The basic unit of the neural network is the perceptron. The inputs are the two notes, n1 and n2, each with its corresponding impact on the final outcome. The 2 outputs will be 1 if it is approved and 0 if you won’t pass the exam.

 How do we use a neural network to find out how much each exam is worth? Here the fundamental unit of the neural network will suffice the perceptron. A perceptron is an element that has several inputs with a certain weight each. If the sum of these inputs for each weight is greater than a certain number, the output of the perceptron is one. If it is less, the output is zero.

 In our example, the inputs would be the two test scores. If the result is one (that is, the sum of the marks by their corresponding weight is greater than five), it is a pass. If it is zero, fail. Weights are what we have to find within the training. In this case, our training will consist of starting with two random weights (for example, 0.5 and 0.5, the same weight for each exam) and seeing what result in the neural network gives for each student. If it fails in any case, we will adjust the weights little by little until everything is well adjusted.

 For example, if a student with a very good grade in the second exam has failed the course, we will lower the weight of the second exam because it clearly does not influence too much. Little by little, we will end up finding the weights that fit the notes that the teacher put. The idea of the adjustment or feedback is to adapt the network to the "hidden" information that the data that we pass to it actually teaches it.

 As we said before, this is a very simple example but we hope that it helps you to understand the whole picture of what a neural network is in a simple example. 

So how else do they work and what are the possible areas of application which are a reality, not a sci-fi dream yet?

The human brain is a model for artificial neural networks. These networks can learn to act independently and thus help to implement automated driving. Artificial neural networks (ANN) simulate information processing based on the model of the human brain. The reason for this is that “conventional computer programs are not well suited for all tasks, while human thinking can solve them without any problems”

Examples of this are pattern recognition, speech recognition, or forecasts.

Accordingly, ANNs are developed on the model of the human brain and trained using a large amount of data. Artificial neural networks aim to simulate the principle of learning in the form of a computer program.

Automated and autonomous driving in SAE Levels 3 to 5 are among the core topics of future mobility. It is extremely important to recognize the surroundings and precisely plan trajectories using artificial intelligence. Machine learning is used to improve localization, networking, and cloud applications.

The learning then works as follows: New skills are learned through imitation using the principle of trial and error. As in the human brain, the neuron and the networking of neurons also play a decisive role in ANNs. The artificial neurons, also known as nodes, are arranged in layers that fulfill different functions. Each layer consists of a large number of individual neurons, the number of which varies depending on the architecture. Each neuron in a layer has an underlying network with the neurons of the next layer. A fully connected layer (FCL) is used when every neuron in one layer is connected to all neurons in the next layer. The intermediate layers are referred to as hidden layers since typically only the first and last layers are addressed for input and output.

 Neural networks are beginning to solve problems that eluded from the computers.

The more data, the better

When learning, the networks change all information or numerical values at the connections between the nodes until the results are good enough. In their inner layers, the networks independently develop compact representations from the raw data, which means that many preprocessing programs are superfluous and the actual task is to make it easier for them to learn. The model developed by the network can then be applied to new, potentially unknown data of the same type.

In general, the following applies: The number of layers determines, among other things, the degree of complexity that an artificial neural network can map. Key knowledge obtained, which have a particularly large number of hidden layers between the input and output layers, make a neural network "deep", which leads to the concept of deep learning. And the more data is available, the better the neural network will work. Big data and artificial intelligence complement each other perfectly: Neural networks need big data to be well trained, and big data require new computer models (especially neural networks) since conventional computers can no longer process the large amounts of data efficiently enough.

 Be that as it may, it is a very interesting field that promises quite a few short-term advances, especially in image and sound recognition. We will be attentive to see what surprises in store for us.