Today many of the advances in the realm of artificial intelligence are new statistical models, but the overwhelming majority of these progressions in technology are called artificial neural networks (ANN).
If one has ever read anything about them before, one might have to read that these ANNs are a very rough model of how the human brain is wired. It is important to take note here because there is a difference between artificial neural networks and neural networks. Although most people drop the artificial for the sake of clarity, the word artificial was initiated to the phrase so that individuals in computational neurobiology can still use the term neural network to refer to their work.
What are Artificial Neural Networks?
To sum, Artificial neural networks can be described as one of the main tools utilized in the field of machine learning. Give, as the “neural” part of the name suggests, these artificial neural networks are brain-inspired systems that are intended to replicate the way that we humans learn. These Neural networks consist of terminals that include input and output layers, which also include a hidden layer of entailing units that transform the input into a product that the output layer can effectively use. Thus, these are excellent tools for locating patterns that are far more complex or numerous for a human programmer to understand and extract while the machine recognizes it.
Where do they come from?
Yes, it’s true that artificial neural networks are not a new concept. In fact, we didn’t even always call them neural networks and these networks certainly don’t look the same now as they did at their earlier. An interesting fact, during the 1960s people, used what was called a perceptron. These Perceptrons were made of McCulloch-Pitts neurons which were very efficient. To add, we even had biased perceptrons, and resultantly people started creating multilayer perceptrons, that are synonymous and interchangeable with the general artificial neural network that we now hear about.
Uses and limitations of Artificial Neural Networks
A popular question pertaining to this subject asks, “what tasks can a neural network not perform?” Yes, many tasks from, to generating shockingly realistic CGI faces, to machine translation, to fraud detection, to reading and speaking our minds, to understanding when a dog is in the garden and turning on the alert system; neural nets are behind many of the biggest advances in the field of A.I.
Speaking in a broader sense, these networks are designed for identifying patterns in data. These could include specific tasks such as classification of data (classifying data sets into predefined classes), clustering of data (analyzing data into different undefined categories), and prediction of the aforementioned data (using prior events to guess future ones, like the stock market or movie box office).
How Artificial Neural Networks work?
The real question lies in everyone’s mind is “How do these networks learn what calculations to perform?” Was I right? The answer to this is that one needs to essentially ask a series of questions, to get the right answer.
Artificial Neural Networks are usually based on a study in the field of supervised learning. With more and enough examples of question-answer pairs, the calculations and values asked and answered are stored at each neuron and synapse is slowly adjusted. This process is called backpropagation.
Types of Artificial Neural Network
Recurrent Neural Networks (RNN)
Recurrent Neural Networks was initially created to address the flaw in artificial neural networks that did not make decisions based on previous knowledge. A typical ANN was compelled to learn to make decisions based on the context of training. Once it was making decisions for use, these decisions were made independent of each other.
Convolutional Neural Networks (CNN)
Convolutional Neural Networks (CNN), often known as LeNets are artificial neural networks that work on a system where the connections among layers appear to be somewhat arbitrary. Although, the reason for these synapses are to be set up the way they are is to help reduce the number of parameters that need to be optimized. This is achieved by noting certain symmetry between how the neurons are connected. Thus here one can essentially “re-use” these neurons to hold identical copies without really needing the same number of synapses.
CNN’s are commonly used in working with images thanks to their ability to identify and locate patterns in surrounding pixels. It is true that there is redundant information contained when you look at an individual pixel compared to its surrounding pixels. Here you can actually compress some of the information thanks to the network’s symmetrical properties.
Reinforcement Learning
Last but not the least, the last ANN type which is going to be discussed is called Reinforcement Learning. Reinforcement Learning is a broad term used for describing the behavior that computers showcase when they are trying to maximize a certain reward. This innately means that it in itself, this ANN isn’t artificial neural network architecture. However, one can apply reinforcement learning or genetic algorithms in order to build an artificial neural network architecture that one might not have used earlier.
Conclusion
To conclude now that you hold the basic understanding of an ANN, it is important to remember that these networks are powering just about everything and anything that we do today. Many subtle yet crucial functions are formed by these ever-growing networks. The future of ANN holds immense potential to expand and reach all sets of data in our lives.
Andrzej Kopera • Sep 05
