Increasingly, technologies tend to combine with each other, reinforcing each other. An AI-enhanced convergence that can change the course of this century for the better.
By Giorgio Metta
The universe of science and technology is in turmoil and evolution. More and more often we read about acronyms like BANG (bits, atoms, neurons, and genes), NBIC (neuro, bio, info, cogno) in the context of trends related to converging technologies. Many experts theorize a convergence of these four technologies towards a "unique" which in the near future will represent all of our "engineering".
Examples of converging technologies are, for example, synthetic biology (which allows cells to perform operations different from those for which they have evolved, including digital computing), neuromorphic technologies which imitate the functioning of neurons in silicon of our brain, the use of DNA as a mechanism for storing information, the development of nanobots for super selective treatment of important pathologies, etc.
All in all, the idea is simple: the ability to manipulate information, to "line up" the atoms one by one, to understand what our neurons do and the understanding at the molecular level of biology will allow a very rapid advance in our abilities to make digital and quantum devices, to heal and improve ourselves, to take care of the our planet and, in a revolutionary way, improve the quality of life. In these pages I would like to give an interpretation of the same technologies and their convergence in the context of the recent very rapid development of artificial intelligence algorithms.
To be precise in the use of terms, artificial intelligence is used in this context in its "widest" sense, including both symbolic and sub-symbolic methods (such as neural networks) and, at the same time, the techniques of optimization but also all the algorithms that "learn" starting from the data. This is clearly a very broad field of research.
The formidable impact of artificial intelligence
The reason I think artificial intelligence is an important element of technological development is because it itself has an iever-deeper impact on the way we do science. Researchers themselves, regardless of their specialty, are and will increasingly have to adapt to working side by side with "algorithms". The data will become essential, its mismanagement will be fatal.
It will be something to be carefully collected, labeled, stored in ways and formats that allow its subsequent algorithmic valorisation. If until a few years ago, the European recommendations regarding FAIR data (Findability/Travabile, Accessibility/Accessibile, Interoperability/Interoperabile, and Reuse/Riusabile) were something abstract for the researcher, today they become a tool to be competitive.
There has recently been talk of the so-called foundational models (Foundation Models), those on which a variety of specialized applications should be based. Patterns typically composed of hundreds of billions of parameters, trained semi-automatically on very large amounts of data, who perform tasks such as conversing in natural language. Well, some researchers have begun to use them to automatically generate codes in the most common programming languages by answering questions in natural language such as: "Would you visualize the data contained in the variable X and Y using a scatterplot".
The programmer asks, the AI prepares the code and adds it to the developing program. Result, the programmer becomes extremely efficient, no longer has to worry about the details that are continuously verified and checked by the AI. It has been estimated – to date – an acceleration of the 30% in the preparation of a code. Imagine the value for companies. Own programmers get the 30% faster on average. This is the deep meaning of modern artificial intelligence engineering.
It is clear that the competition is only for those who can access and pay for AI software of this type. We reason - as a country - whether a very important investment is not necessary to build a very high competence in the AI domain.
New materials
A few years ago, the scientific community began studying graph neural networks to represent and learn the possible synthesis paths of new materials. Chemistry has been encoded simply from step sequence data and from a set of reactants, catalysts, physical parameters, etc. leads to the synthesis of a certain product in certain quantities. The latest results have shown that these neural networks are about 10% more efficient and reliable than a “human” chemist. They certainly don't invent anything new in the strict sense but they do very well in a profession which would in any case involve a considerable expenditure of time for the human being.
Imagine labeling everything that is done in the laboratory (or in many laboratories) for several years and training increasingly sophisticated models, then linking them to basic models and knowledge of physics rather than what is already known to chemistry as a whole. We can hypothesize very rapid development processes for new materials. Since it is subsequently necessary to try what the AI suggests and, probably, also to refine the boundary conditions to obtain the required accuracy, one can subsequently imagine robotizing the laboratory.
Again, converging technologies – physics, chemistry, artificial intelligence – are making research faster, allowing us to explore new solutions better than we could manually. It is natural to wonder if it will still be possible to do research in the world of materials without having an AI accompanying us.
Think of how important these new materials are in medicine for making increasingly effective drugs, in the fight against climate change for increasingly high-performance solar panels, enzymes for "digesting" plastic, but also for capturing CO2 excess in industrial processes.
The revolution in medicine
About 1 billion human beings are believed to suffer from mental disorders. The 70% of these are linked to depression and anxiety states. Let's talk about one real health emergency that will affect our society in parallel with an aging population that sees the number of individuals over 65 reach more than 20% of the total world population by 2050.
Important health problems accompany aging with neurodegenerative pathologies. It is absolutely necessary to understand what the causes are by observing and modeling the functioning of the brain. The ability to observe the brain with unprecedented resolution is provided by electrodes with long-term stable implants, among other technologies.
It is hypothesized to be able to reach hundreds of thousands of neurons, collect signals which can then be modeled using artificial neural networks, extract their meaning with respect to behavior, distinguishing anomalous elements. A future where recording and stimulating neurons will make it possible to determine causes and remedy them, perhaps by reconstructing neural pathways that no longer function. The amount and complexity of information can only be analyzed through AI.
Speaking of medicine, in the broadest sense of the term, AI is capable of doing a great job of modeling drug-receptor interactions rather than resolve and reproduce the molecular structures of the cell such as, for example, proteins.
Many recent studies have begun to study them using machine learning approaches revolutionizing research in the life sciences. Electron microscopes allow us to see the structure of proteins, optical ones to observe them inside cells while they carry out their function, sequencers to record the structure of nucleic acids and their modifications.
They are all immense data sources for which it is certainly not possible to hypothesize a manual analysis. Bioinformatics once again mixes knowledge. More and more, however, important elements of this discipline are implemented through AI algorithms. From silicon we then move on to chemistry and the solutions are tested first in vitro to then arrive after a long journey to clinical trials on human beings.
There are about a thousand genes that encode proteins for which we can inhibit their expression through drugs. Still few compared to the approximately 22,000 that we find in our DNA in addition to the approximately 60,000 non-coding ones. So imagine what it might mean to understand from in silico analysis how to intervene on a specific cellular process by increasing or reducing the production of a certain protein. We could cure diseases that are totally incurable today. Bring relief to those who do not currently have it. Again, converging technologies can have incredible results.
Ultimately, the growth of AI is certainly not limited to its more evident applications such as those of the automatic control of increasingly autonomous robots, means of transport that actually become robots rather than their use in information systems as such. In the world of increasingly converging technologies, AI certainly becomes the new electricity, the force that can change the course of this century for the better. Given the problems we will have to face, one above all climate change, we can only hope that there will be more and more artificial intelligence alongside the human one.
Giorgio Metta he is Scientific Director of the Italian Institute of Technology (IIT).
