The Future of Artificial Intelligence

In recent years, artificial intelligence has appeared widely in the real world as well as in science fiction films. Navigation apps, language-to-language translation, voice search have become part of our daily lives. We come across mobile robots, self-driving cars, news about software that diagnoses cancer in the media almost every day. Although artificial intelligence has partially proven its maturity thanks to deep learning methods, which were popular in the 2010s, its impact on the world is still limited. AI is very successful in clearly drawn areas such as playing chess, solving mathematical problems, playing computer games, recognizing and naming objects. On the other hand, artificial intelligence systems turn into fish out of the water as soon as they enter our world.

Although it has some weaknesses, the human brain remains the most wonderful object we know in the universe. Artificial intelligence systems cannot yet be an alternative to the human mind for the reasons I will explain below.

Energy Consumption

Machine learning models, the main component of Modern artificial intelligence systems, are learning through trial and error. This requires hundreds of thousands of transactions over huge mountains of data. Computer processors' performance has improved geometrically over the years, but processors' energy efficiency has increased arithmetic. This leads to a rapid increase in the energy costs of artificial intelligence systems.

According to Ray Kurzweil's 1999 estimate, the human brain can perform 2x10^16 operations per second. These days, we finally live in the times when the most powerful supercomputers in the world reach this speed. However, it is enough for us to eat two eggs for breakfast to work, while these systems, consisting of tens of thousands of processors, consume the electricity for a town. Supercomputers are currently used for weather forecasting, oil exploration, molecular modeling, and astronomical data processing.

Algorithms

Despite intensive research, it is not understood how the human brain works. Neurologists know that the brain has a network structure, performs its internal communication through electrochemical processes; some of its regions take on functions such as speech, vision, memory. Although what is known about the brain has grown like an avalanche over the years, we still do not have a blueprint showing the brain's work principles. It is assumed that deep artificial neural networks, which make artificial intelligence popular, work similarly to the brain.

It takes a huge pool of data to train deep artificial neural networks. But using them, we can create models that can separate the cat from the dog, the sound of the wave from the sound of the wind. However, a child does not need to see millions of animals to get to know animals. When we show a single cat and say, ‘this is a cat,’ he can understand what the cat looks like.

Another weakness of Deep Artificial Neural Networks is that they cannot use what they have learned in another context. A model that we train on a particular topic cannot transfer the information it receives to models on similar topics.

Purpose and Emotions

People are usually chasing one or more targets at a given moment. Even a person who has no life goals has to satisfy his needs, such as sleep, hunger, thirst, to survive. Creating an algorithm that can set its own goal is considered the holy grail of artificial intelligence work. For example, some companies want to learn the prioritized work they need to do from artificial intelligence by giving all kinds of data in their databases. However, artificial intelligence algorithms need a pointed goal to produce a meaningful result in their present state.

Emotions play as much a role as rational data in the decisions people make. For artificial intelligence to be active on earth, it will also have to have emotional intelligence. Although there are studies to determine people's mood from facial expression and tone of voice, more is certainly needed for effective communication.

Although planes look quite like birds, they don't flap their wings. We can assume that general-purpose artificial intelligence will have sides that look and don't look like humans.

Popular Research Topics

Artificial intelligence's weaknesses are solved by new methods being explored around the world.

Dueling Networks (Generative Adversarial Networks) has recently come to prominence as a key model for producing realistic images, sounds, and text. Dueling Networks contain two artificial neural networks with opposite goals. One network (manufacturer) learns to produce realistic photos from input data. In contrast, the other network (distinctive) evaluates the first network's output by determining which photos are real and fake. Using networks of contention, it is possible to produce celebrities who do not actually exist, compositions similar to Bach compositions, and paintings in Van Gogh's style. I want to note that these models are an important step towards artificial intelligence demonstrating creativity.

Capsule Networks architecture emerged as a major software innovation in deep learning. Traditional deep learning techniques were successful in image classification and tagging. However, they do not learn spatial relationships between objects (for example, face and mouth, nose and eyes, and the object's relative positions when viewed from different angles. Capsule Networks have an advanced hierarchical structure that is supposed to resemble the way humans process information. Traditional artificial neural networks contain multiple layers, and layers of a neural network are described as a nested Collection capsule: Each capsule contains a small group of neurons in each layer of multiple capsules. A capsule serves as a special software module that learns to detect a specific pattern or object in an image.

One-Shot Learning, or learning with little data, has recently come to prominence as a paradigm aimed at faster and more efficient learning. Such an artificial intelligence system is intended to classify data using a tiny training set. We can say that the design that makes learning models possible with a small dataset is the transfer of information from previously learned categories to new categories. Studies are being conducted on the creative use of models such as Bayesian networks (Ing: Bayesian Networks) and artificial neural networks supported by short-term memory to enable information transfer.

Result

There is a consensus that artificial intelligence systems will play an important role in the world's future. However, whether they will one day exhibit a cognitive performance close to humans remains a matter of considerable debate. My personal opinion is that in the coming years, as today, there will be jobs that people and machines do better, but the number of jobs that machines do better will increase by the day.

I am very interested in what this new type of consciousness, which is sprouting before our eyes, will turn into, for example, 10 years from now. It seems that we will continue to discuss the issue of artificial intelligence for longer.

Image Sources: pixabay.com and bilimkurgukulubu.com