Artificial Intelligence: a look at neural networks

One of the most complex and misunderstood topics that make headlines lately is artificial intelligence. People like Elon Musk warn that robots could one day destroy us all, while other experts say that we are on the verge of an AI winter and that technology is not going anywhere.

Artificial intelligence has become a focal point for the global technological community thanks to the increase of deep learning. The radical advance of artificial vision and the processing of natural language, two of the most important and useful functions of AI, are directly related to the creation of artificial neural networks.

Scientists believe that the brain of a living creature processes information through the use of a biological neural network. The human brain has up to 100 trillion synapses (gaps between neurons) that form specific patterns when they are activated. When a person thinks of something specific, remembers something or experiences something through their senses, it is believed that specific neuronal patterns "catch" inside the brain.

In the field of deep learning, a neural network is represented by a series of layers that function similarly to the synapses of a living brain. We know that researchers teach computers how to understand what a cat is, or at least what the image of a cat is, feeding it with as many cat images as they can.

The neural network takes those images and tries to discover all the similarities between them, learning this you can find cats in other images.

Scientists use neural networks to teach computers to do things on their own and solve a wide variety of problems. To understand a little how they work, and how computers learn, let's take a quick look at three basic types of neural networks:

Adverse generative networks

Ian Goodfellow, one of Google's AI gurus, invented the adverse generative networks (GAN - generative adversarial network) in 2014. A GAN is a neural network composed of two opposing sides, a generator and an adversary, who struggle between them until the generator wins. If you wanted to create an AI that imitates an art style, like Pablo Picasso's, for example, you could feed a GAN with a lot of his paintings.

Adverse generative networks are used in a wide variety of AIs, including one built by Nvidia that creates "real" photographs of people who do not even exist.

Convolutional neuronal networks

Convolutional neural networks (CNN) have existed since the 1940s, but thanks to the development of more efficient hardware and algorithms, they are now becoming useful. A CNN has several layers through which the data is filtered in categories. This type of neural networks is used mainly in the recognition of images and the processing of the text language.

You could train a convolutional network by feeding it with complex images that have been labeled by humans. IA learns to recognize things like stop signs, cars, trees and butterflies looking at photos that humans have classified, comparing the pixels in the image with the labels they understand and then organizing everything they see in the categories in which they have been trained .

The CNN are among the most common and robust neural networks, even used in the diagnosis of some diseases, where they have been able to overcome the performance of real doctors.

Recurrent neural networks

Finally we have the RNN (recurrent neural network), or recurrent neural networks. The RNN are used mainly for the AI ​​that requires nuances and context to understand its input.

An example of such a neural network is an AI of natural language processing that interprets human speech. Examples that many people know are found in the Google Assistant, and Alexa, Amazon, which are cases of RNN for everyday use.

There are more types of neural networks, only three of them have been mentioned and quite simply. If you want to expand your knowledge on the subject, we suggest this free course on Artificial Intelligence prepared by the University of Helsinki and Reaktor, among many other resources that are available on the web.