I am fascinated by the human capacity to create models and architectures of the most creative to solve problems or generate powerful products using artificial intelligence. The use of hybrid architectures makes it possible for smart devices like Sophia, from Hanson Robotics. It has its architecture developed by Ben Goertzel and his team, who are now at the head of SingularityNet. Sophia combines robotics for facial movements, image recognition, natural language generation and processing, just to name a few that can be said, without knowing more details about its architecture.

In January of this year, Samsung presented at CES 2020 its version of artificial humans, called NEON. According to the manufacturer, it should be available for beta testing by the end of the year in some locations. I had already talked about digital humans in my last article "Artificial Intelligence: technology that you can't help but understand". Artificial humans are examples of combined architectures that can mimic a part of the activities that we humans can do, such as interacting, moving, talking and answering questions. The result is impressive.

One of the most potential fields for creative architectures within artificial intelligence is deep learning. In a very simplified way, deep learning uses neural networks, algorithms that are based on the human brain with mathematical functions that simulate a biological neuron, with many layers, in order to combine data, identify patterns and learn through them.

An architecture now well known to most people who work with machine learning is called GAN –Generative Adversarial Networks. The GAN architecture is based on two neural networks with different “skills” linked together.

The first of these networks is a neural network programmed to generate images of a certain nature. These images can be totally random, or already have a pre-established pattern coming from previous “trainings”, such as images of people, images of cars, images of rooms at home ... The second network is a discriminatory network, it discriminates images values ​​of a given pattern: images of people, cars and rooms of the house, in relation to false images or that do not represent that pattern. She was "trained" only by real images.

In the GAN architecture, you connect one neural network to another. The objective of the image-generating neural network is to be able to “trick” the discriminating network by making this network believe that the artificially generated image is a true image.

The architecture is built in such a way that every time the image generating network “deceives” the discriminating network and generates an image with such good quality - that it is considered a real image of a person, car or room - it gains a kind bonus (her artificial synaptic weights are increased), but every time she does not achieve this feat, she has a kind of punishment (her artificial synaptic weights are reduced). This creates an image generating network that converges to become very efficient in generating artificial images as good as the real ones.

One of the examples that became well known to the world was “This person does not exist”, or in Portuguese “This person does not exist” - a website that each time it is accessed, creates an image of an artificially generated person. A person who never existed on the face of the Earth, but who has an image so rich in details and so similar to a human that there is no evidence that we can present that it is an unreal image. And just to make the difference clear, it has been a long time since we have had technologies to create very good artificial images like the real ones. The big question here is that a few years ago we created an algorithm that generates real images instantly and of the highest quality. We created the machine to generate these images.

GAN architecture has revolutionized the world of deep learning and is also one of the most used architectures in what in popular jargon is known as “deep fake”, the construction of news and false images using artificial intelligence and deep learning. In other words, there are still many ethical and governance issues that we need to discuss before using this powerful technology, and that will be the subject of another article on the topic.

There are numerous interesting neural network architecture applications. Another one that I really like is an architecture known as “Hierarchical Deep Models”. It is another way of organizing neural networks in order to be able to quickly understand low-level patterns such as shapes and colors, in order to, from this combination, generate new elements based on very little training data.

There are examples that generate sketches of images like fruits and animals that are based on just three models. That is, here too, there is an emphasis on the way in which neural networks were organized: (1) first one learns through simpler neural networks to extract the basic patterns of an image, (2) then networks are organized higher level that connect these networks of basic standards, creating a correlation between these different standards, and (3) for each new category to be trained, a new hierarchy and priority relationship is created between the most basic networks.

This architecture generates very efficient learning structures that they can understand using very few examples, and bring impressive results to algorithms that basically use mathematical functions.

There is a fantastic field for the construction of new architectural models and new solutions to various problems that we face in our daily lives. Human imagination combined with knowledge of technologies related to artificial intelligence, without a doubt, has revolutionary potential in all aspects.

It is not for nothing that I am so enthusiastic about this technology, I believe that we have a lot to do to transform the world we live in into a much better world. Artificial intelligence is undoubtedly a technology that can greatly expand the entire human potential, in this sense. Perhaps you are one of those people who build something revolutionary?