«Breast cancer has had less impact due to covid than other oncological pathologies»
9 March, 2021
Women and people with mental illnesses: a double stigma
9 March, 2021
Current computing power already allows artificial intelligence, through deep learning, to establish scientific laws based on experimental data taken from reality. Will artificial intelligence make the science of tomorrow? And will we even know how to understand it?
The novel 'Do Androids Dream of Electric Sheep?' It is one of the best known within the science fiction genre. It gave rise to a very famous film adaptation, 'Blade Runner', and is about the blurred boundaries between artificial and human intelligence. We witness astonished, almost every day, the samples of what artificial intelligence is capable of doing, we carry one or more in our pocket, in the form of a mobile phone, and we have seen robots beat human champions at chess.
However, deep down we harbor some hope that artificial intelligence will not be able to completely replace us, clinging to the most human activities: dreams, creativity or, why not, science. Can an artificial intelligence do science, create beautiful scientific theories such as general relativity, universal gravitation or quantum chromodynamics, to give just a few examples? Doesn't the development of a new scientific theory require creativity? Do robots have that?
We will see if it does. In the Artificial Intelligence laboratory of the Engineering Research Institute of Aragon (I3A), several groups of researchers are dedicated to developing the ingredients that make it possible, within an international trend that, logically, has aroused great interest.
But let's go by parts.
What does artificial intelligence know how to do?
In reality, an artificial intelligence knows how to do very few things. However, skillfully combined, they give rise to an unthinkable number of possibilities. Within machine learning techniques, three types of tasks are usually distinguished: supervised learning – that in which the machine is 'trained' by providing it with a certain number of data, along with the conclusion that the machine must draw in this regard; unsupervised learning – which allows machines to do things like classify data into similar groups or know which of them are important and which are not; and reinforcement learning – that in which the machine is provided with certain incentives to learn, in the same way that we give a horse a sugar as a reward.
Of all of them, the one that interests us most now is supervised learning and, particularly, regression. In statistics, the process by which relationships between different variables are determined is called regression. Well, the process by which an artificial intelligence establishes a scientific law is neither more nor less than a great regression, surely truffled with other unsupervised learning techniques.
So, if regression has been known for centuries, why are we now talking about scientific artificial intelligence? Although the first person to think about the possibility of certain machines carrying out reasoning was the Mallorcan Ramón Llul, in 1315, artificial intelligence has gone through a number of what are known as winters of artificial intelligence, times in which Scientists in particular and the public in general fall into a general pessimism about its usefulness and, therefore, about the convenience of financing this type of research. We have emerged from last winter in part thanks to the popularization of 'deep learning', something that was already known, but which has triumphed in an unprecedented way thanks to the exponential growth in the capacity of computers. Deep learning is capable of recognizing animals in photographs, understanding natural language or piloting airplanes or cars. This computing power has also allowed deep learning to establish scientific laws from experimental data about the reality around us.
