The Authors
Metta
Long a coordinator for the development of the iCub robot, in fact, the reference platform for research in AI, he focuses his research activities in the field of bioinspired systems and humanoid robotics with particular reference to the design of machines capable of learning from experience.
The new electricity
“Artificial Intelligence is the new electricity.” These words, uttered by American "colleague", computer scientist and Stanford University professor Andrew Ng, were intended to emphasise the change generated by AI today, similar and equivalent to the impact that electricity had in the 20th century. I would like to provide within this same vision an only slightly different interpretation. The use of artificial intelligence is so pervasive that it reaches every segment of our daily life, which is thus affected and changed: health, agriculture, transport, and energy are some of the sectors that will be intertwined with AI. As electricity did a hundred years ago, AI will become the necessary and indispensable condition for a higher quality of life.
In the 1950s, Alan Turing formalised the concept of computation using very simple elements, precisely defining the concepts of algorithm and digital computation. The English mathematician’s incredible curiosity also led him to wonder whether a machine could actually think. Almost a century later, the brilliance of his insights is the theoretical basis on which the new technological revolution is founded. Indeed, the last fifteen years have been fascinating in terms of progress thanks to the growth of computing and information storage capacities, aided by a vast and widespread Internet network, which in its development collects data on decades of human activity in any field. It is here, within a network filled by our everyday life also and above all through mobile devices, that AI developers have found an easily accessible and almost unlimited database from which to train models with trillions of parameters.
It is therefore even more evident that the world in which AI moves and, metaphorically, lives today is the completely digital world of cyberspace. Not for nothing do the best results come from "symbolic" processing, in which the alphabet is identifiable with absolute precision. Worlds where a "language" - not necessarily the spoken one - implies the structure of what is attempted to be represented in the algorithm. Human textual language, such as programming language, or even the language of the three-dimensional structure of proteins, or energy or transport networks, are all united by graph structures, i.e. geometric figures made up of a finite set of points, whose nodes - or vertices as you like - are words, symbols, chemical interactions or state transitions. This proximity between symbolic data and algorithm has led to the prevalent use of AI within digital devices. It is no coincidence that the first field of application is our mobile device, as well as our computer and interfaces.
However, there is one territory of AI application that is still very difficult: the "physical", real world. The interaction between the paradigm of digital computing and that of the real world, which is analogue by definition and, above all, unpredictable also because it is populated by us humans, is still very difficult for artificial intelligence. The difficulty is almost epistemic. Algorithms must extract the symbols themselves in order to process them and make accurate decisions so that movement control is error-free, safe and ultimately effective. The algorithms we use are not yet precise enough.
Hence the broken promises. The self-driving car is perhaps the best example. It is very difficult not to get stuck because of an unforeseen event: it snows during the night and in the morning my driverless car gets stuck because the landscape is visually different from what had been assumed.
To progress in this field, we therefore study the interaction between man and machine as an epitome of the relationship between analogue and digital. Robotics - usually representing the mechanical and the exact - takes on the role of the interface between us and cyberspace. Here, in this place between digital and analogue, the AI algorithms live. They control the robot’s movements as they interact with us. What is fascinating about this research, where experimental psychology and engineering meet, is in fact how it enables both improved movement and performance of robots and a better understanding of human beings. We find ourselves studying how different robot behaviours become communication signals and how these influence our perception of them and consequently our responses. In addition to becoming the means to improve the interface between digital systems and human beings, robots - in these typically humanoid experiments - teach us something about cognitive functions: attention, the ability to prehend and manipulate objects or, for example, non-verbal communication itself. In this field, our research leads us to hypothesise that in order to really learn to reason like us humans, the newly constructed magnificent artificial neural networks must also be equipped with a body similar to ours.
In more pragmatic research that is close to application, machine learning (one of the AI technologies) has been successfully applied to improve simulations of new materials and understand how they work, the synthesis of the same materials in chemistry laboratories, the interpretation of genomics data but also the prediction of the three-dimensional structure of proteins. Materials developed in this way find applications in the field of energy, improve the functioning of solar panels, the efficiency of turbines, the precision of medicine through the understanding of biology. AI is becoming an important aid in solving the many technological problems that enable us to take care of our planet (especially new materials) and ourselves (health). Artificial intelligence has become an incredible accelerator in the way we construct experiments, how we conduct them - sometimes even with robots - and how we design future experiments. It is clear how scientists or companies with access to algorithms, data and computing power will be able to compete at lower costs while generating results faster and faster. In my view, AI is much more than electricity. It is not just technology, but something that changes the very way we conduct science and develop other new technology. Generative models are drawing a new scientist who conducts experiments sparingly, because most of them are already simulated inside a high-performance computer.
We can wonder what the near future will look like. I see two challenges. The first is the technical one. Taking AI technology to such a sophisticated level of development that it becomes the intelligence that helps us solve sustainability and health problems. While today we can use it for a few new experiments, in the near future it will enable us to analyse our scientific knowledge in a broad sense, linking the results of millions of experiments and interpreting them systematically.
The second challenge is to dream big and use AI as a tool to really understand what our brains are like, on the one hand to cure them, and on the other hand because understanding ourselves to the core has been the dream of mankind ever since it first raised its eyes to the sky and observed the space around it.