{"id":4547,"date":"2024-10-06T17:19:14","date_gmt":"2024-10-06T15:19:14","guid":{"rendered":"https:\/\/www.rivistaeco.com\/?p=4547"},"modified":"2024-10-06T17:19:14","modified_gmt":"2024-10-06T15:19:14","slug":"artificial-intelligence-supporting-science","status":"publish","type":"post","link":"https:\/\/www.rivistaeco.com\/en\/2024\/10\/06\/artificial-intelligence-supporting-science\/","title":{"rendered":"Artificial Intelligence Supporting Science"},"content":{"rendered":"

If everything can be explained by statistical correlations, what is the role today of the scientific method that has guided the development of our knowledge since Galileo? For now, it remains the only method capable of constructing interpretable, controllable theories adaptable to unseen situations. AI is a valuable aid in finding hidden links among a vast number of variables.<\/span><\/i><\/p>\n

\u201cThe end of theory: the data deluge makes the scientific method obsolete.\u201d This was the provocative title of an article by journalist Chris Anderson published in 2008 in <\/span>Wired Magazine<\/span><\/i> (of which he was then editor-in-chief). The latest developments in artificial intelligence rekindle this thesis and make the arguments I used five years ago to refute it, in an issue of <\/span>Le D\u00e9bat<\/span><\/i> dedicated to the question tormenting researchers, relevant again: are we at the end of the scientific method?<\/span><\/p>\n

The Intellectual Revolution<\/b><\/h3>\n

Let’s start with a fact: the information revolution (methods of writing, storing, sharing, communicating, social networks, and now artificial intelligence) represents a qualitative leap. Unlike the great transformations of the past that changed our access to food or energy (the Neolithic agricultural revolution ten thousand years ago or the industrial revolution one hundred and fifty years ago), today we are facing an “intellectual” revolution that affects the immaterial and changes the way we use our brains, disrupting the organisation of our society whose economy is based on the tertiary sector and thus on the (more or less sophisticated) processing of information. We can therefore expect the new revolution to transform humans not so much biologically but in the way they use their resources. Michel Serres’ “Petite Poucette” (born with a smartphone and skilled at typing with her two thumbs) always has an encyclopedia, various dictionaries, and large parts of our libraries at her disposal in her pocket. With increasingly sophisticated programs capable of classifying and retrieving information, access to personal memory and a large part of universal knowledge is multiplied. In this sense, we are experiencing a new phase of brain “liberation”, similar to the inventions of writing or printing. Even in 2008\u2014before the arrival of the “new AI”, born in 2012 with “deep network” algorithms that marked a decisive turning point in image analysis and classification\u2014Chris Anderson warned us: “With big data, the [old] approach to science\u2014formulating a hypothesis, then a model, then testing\u2014is becoming obsolete; (…) Now there’s a better approach. Petabytes allow us to say, ‘correlations are enough’.” From Galileo onwards, we had learned to formulate hypotheses, build models, and conduct experiments to confirm or modify them. Now, however, the entire apparatus of the scientific method becomes superfluous: results can be obtained without going through models merely by relying on an immense number of correlation calculations in enormous databases. Indeed, at first glance, the recent successes of the new AI seem to support Anderson’s prophetic text, but it is necessary to delve into the details to make this judgment more nuanced.<\/span><\/p>\n

At the Heart of Everything is Machine Learning<\/b><\/h3>\n

Modern AI is based on an idea\u2014”machine learning”\u2014that dates back several decades, but has only recently found application thanks to the arrival of powerful computing resources associated with enormous databases. Let’s take image analysis. A three-year-old child can tell if a photo has a cat or a dog in it. It may not seem like much, but this simple task represented a challenge for computers that experts wrestled with. Now the problem is solved because we have machines (actually software programs, i.e., algorithms) capable of distinguishing cats from dogs. In fact, they can do much more: they can identify faces and, from a simple photograph of a skin lesion, they can suggest a preliminary diagnosis indicating whether it is probably benign or poses a cancer risk, with a diagnosis as reliable as that of the best dermatologists. The basis of this decisive progress is a simple idea: instead of trying to program the computer, let it program itself during a learning phase by adapting its structure (in reality, the billions of parameters that define its structure) based on examples. For this, therefore, it is necessary to have a database containing hundreds of thousands of examples of animals, with each image annotated: so that its content (cat or dog) is known. The program will choose the parameter combination, from the immense number of possible ones, that makes the fewest possible errors during the entire learning process. This process represents a complete paradigm shift because it is the machine that learns to program itself, based on an enormous number of examples. There are indeed two great surprises of the last ten years: the first is that learning is possible (even though the number of possible parameter combinations in the machine far exceeds the number of atoms in the universe); the second is that the machine thus trained can generalise: if presented with a new image of a cat, even if it has never seen it before, it will identify the presence of the feline. All recent AI advances are founded on this basis, and now go far beyond image analysis since we can process and generate text (the famous ChatGPT), images, and videos, with AI called upon to contribute to almost all scientific fields whilst also gaining a foothold in businesses and administrations. Chris Anderson is therefore right on one point: the new machines, after extensive training, detect subtle correlations that we could not otherwise have found and which prove to be extremely valuable. Let’s take one of the most important results in science: protein folding.\u00a0<\/span><\/p>\n

Proteins are built from chains of amino acids which fold to form their 3-D structures.\u00a0 Thanks to AI, we can take an initial chain and predict how it will fold, something which is crucial for understanding its function since it is through the elements on the resultant surfaces that it will interact with other molecules.<\/span><\/p>\n

This is a first step in decoding the great book of biology: by reading the genetic sequence text, we have made a key advance towards understanding its function. The best physicochemical methodologies of the past, developed using the interaction energies between patiently determined amino acids and molecular dynamics or Markov chain simulations, have now been surpassed by a purely statistical approach. A similar surprise came with language models: ChatGPT demonstrates that it is possible to generate syntactically and semantically correct sentences by training a neural network to guess one word after another, based solely on correlations with previous words assimilated in a massive learning phase involving a huge (and poorly controlled) corpus of texts. It is a triumph of the statistical method which\u2014let’s admit it\u2014has proved far superior to anything we could have imagined twenty years ago. But what do algorithms learn? What are their limits and how well do we understand them? First of all, it is essential to realise that their behaviour has nothing to do with that of the human brain. Cognitive science experiments on language learning show us that a small child can generalise a concept from very few examples. Nothing to do with deep neural networks. At the moment, such networks have no representation of the world, neither of its spatial organisation nor of its causal relationships: they can tell you if a cat is present in an image but do not know what a cat is, nor that it jumps, purrs, hunts, and eats. Similarly, a language generative model produces sentences (the most probable word sequences) that make no sense to it. There is no intentionality in its text, even if we readers tend to attribute it, accustomed as we are to texts that until now had always been produced by human beings endowed with intention. A fundamental issue is the interpretability of results. We observe a well-trained neural network on a specific task. We know everything about the process; we have access to the values of the billions of parameters found by the machine during its learning phase, yet we understand very little about how it generalises. When questioned, the algorithm will propose a statistical answer but it is without the ability to provide an argument or a deduction. For now, this is an essential limit: without deduction, completely wrong answers, which cause the biggest problems in applying the new AI to produce an autonomous decision, cannot be avoided.<\/span><\/p>\n

The Importance of the Scientific Method<\/b><\/h3>\n

Modelling in science proceeds in a radically different way. It goes through phases of simplification: identifying relevant variables and finding the laws that connect them. For a complex problem, one arrives at a hierarchy of variables and laws on different scales, whose composition establishes a modelling of reality in the form of a theory with causes and effects and therefore a predictive system. A set of relatively simple laws can lead to very sophisticated theories: for example, the standard model of interactions between elementary particles has rules that can be written compactly on half a page but whose meaning becomes clear only after studying entire books of field and particle theory over five or six years of physics studies. So far, the modelling approach has been the only one capable of constructing interpretable, controllable, and adaptable theories for unseen situations. The correlations found by modern AI cannot replace them, even if they prove very valuable in all phases of modelling a complex system, for example, to identify, among thousands of possible variables, which combinations are important. Therefore, in science, we must see modern AI as a new phase of the digital approach, providing very powerful tools to support the scientific method. At the same time, many scientists are engaged in studying deep neural networks using scientific intelligence in the hope of one day understanding how they work. It will therefore be necessary to get them out of the impasse they have fallen into under the pretext that “everything is in statistical correlations”; because Chris Anderson’s ideological position also reflects a practice. In fact, for now, much of the research is purely empirical, and too little effort is devoted to reintroducing fundamental notions of space, objects, causality, and rules into deep neural networks\u2014in a word, to reintroduce models.<\/span><\/p>\n

Bio<\/b><\/em><\/p>\n

Marc M\u00e9zard is a professor of theoretical physics at the Department of Computing Sciences at Bocconi University. His current research focuses on statistical physics with applications in machine learning, information theory, and computational complexity.<\/span><\/em><\/p>\n","protected":false},"excerpt":{"rendered":"

If everything can be explained by statistical correlations, what is the role today of the scientific method that has guided the development of our knowledge […]<\/p>\n","protected":false},"author":7973,"featured_media":0,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"_jetpack_memberships_contains_paid_content":false,"footnotes":""},"categories":[1],"tags":[],"coauthors":[167],"class_list":["post-4547","post","type-post","status-publish","format-standard","hentry","category-non-categorizzato"],"acf":[],"yoast_head":"\nArtificial Intelligence Supporting Science - Rivista Eco<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"http:\/\/www.rivistaeco.com\/en\/2024\/10\/06\/artificial-intelligence-supporting-science\/\" \/>\n<meta property=\"og:locale\" content=\"en_US\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"Artificial Intelligence Supporting Science - Rivista Eco\" \/>\n<meta property=\"og:description\" content=\"If everything can be explained by statistical correlations, what is the role today of the scientific method that has guided the development of our knowledge […]\" \/>\n<meta property=\"og:url\" content=\"http:\/\/www.rivistaeco.com\/en\/2024\/10\/06\/artificial-intelligence-supporting-science\/\" \/>\n<meta property=\"og:site_name\" content=\"Rivista Eco\" \/>\n<meta property=\"article:published_time\" content=\"2024-10-06T15:19:14+00:00\" \/>\n<meta name=\"author\" content=\"Marc M\u00e9zard\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:label1\" content=\"Written by\" \/>\n\t<meta name=\"twitter:data1\" content=\"Marc M\u00e9zard\" \/>\n\t<meta name=\"twitter:label2\" content=\"Est. reading time\" \/>\n\t<meta name=\"twitter:data2\" content=\"10 minutes\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\\\/\\\/schema.org\",\"@graph\":[{\"@type\":\"Article\",\"@id\":\"http:\\\/\\\/www.rivistaeco.com\\\/en\\\/2024\\\/10\\\/06\\\/artificial-intelligence-supporting-science\\\/#article\",\"isPartOf\":{\"@id\":\"http:\\\/\\\/www.rivistaeco.com\\\/en\\\/2024\\\/10\\\/06\\\/artificial-intelligence-supporting-science\\\/\"},\"author\":{\"name\":\"Marc M\u00e9zard\",\"@id\":\"https:\\\/\\\/www.rivistaeco.com\\\/#\\\/schema\\\/person\\\/3e6fc92658259155b69a6065831f8e9e\"},\"headline\":\"Artificial Intelligence Supporting Science\",\"datePublished\":\"2024-10-06T15:19:14+00:00\",\"mainEntityOfPage\":{\"@id\":\"http:\\\/\\\/www.rivistaeco.com\\\/en\\\/2024\\\/10\\\/06\\\/artificial-intelligence-supporting-science\\\/\"},\"wordCount\":1736,\"articleSection\":[\"Non categorizzato\"],\"inLanguage\":\"en-US\"},{\"@type\":\"WebPage\",\"@id\":\"http:\\\/\\\/www.rivistaeco.com\\\/en\\\/2024\\\/10\\\/06\\\/artificial-intelligence-supporting-science\\\/\",\"url\":\"http:\\\/\\\/www.rivistaeco.com\\\/en\\\/2024\\\/10\\\/06\\\/artificial-intelligence-supporting-science\\\/\",\"name\":\"Artificial Intelligence Supporting Science - Rivista Eco\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/www.rivistaeco.com\\\/#website\"},\"datePublished\":\"2024-10-06T15:19:14+00:00\",\"author\":{\"@id\":\"https:\\\/\\\/www.rivistaeco.com\\\/#\\\/schema\\\/person\\\/3e6fc92658259155b69a6065831f8e9e\"},\"breadcrumb\":{\"@id\":\"http:\\\/\\\/www.rivistaeco.com\\\/en\\\/2024\\\/10\\\/06\\\/artificial-intelligence-supporting-science\\\/#breadcrumb\"},\"inLanguage\":\"en-US\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"http:\\\/\\\/www.rivistaeco.com\\\/en\\\/2024\\\/10\\\/06\\\/artificial-intelligence-supporting-science\\\/\"]}]},{\"@type\":\"BreadcrumbList\",\"@id\":\"http:\\\/\\\/www.rivistaeco.com\\\/en\\\/2024\\\/10\\\/06\\\/artificial-intelligence-supporting-science\\\/#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"Home\",\"item\":\"http:\\\/\\\/www.rivistaeco.com\\\/en\\\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"Artificial Intelligence Supporting Science\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\\\/\\\/www.rivistaeco.com\\\/#website\",\"url\":\"https:\\\/\\\/www.rivistaeco.com\\\/\",\"name\":\"Rivista Eco\",\"description\":\"\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\\\/\\\/www.rivistaeco.com\\\/?s={search_term_string}\"},\"query-input\":{\"@type\":\"PropertyValueSpecification\",\"valueRequired\":true,\"valueName\":\"search_term_string\"}}],\"inLanguage\":\"en-US\"},{\"@type\":\"Person\",\"@id\":\"https:\\\/\\\/www.rivistaeco.com\\\/#\\\/schema\\\/person\\\/3e6fc92658259155b69a6065831f8e9e\",\"name\":\"Marc M\u00e9zard\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"en-US\",\"@id\":\"https:\\\/\\\/secure.gravatar.com\\\/avatar\\\/5c12a7af97e154698492fa5037e6d1df39c941aa1d911976709d2c8342253e33?s=96&d=mm&r=g9af4c0f16d62cde21b270763cad3dba3\",\"url\":\"https:\\\/\\\/secure.gravatar.com\\\/avatar\\\/5c12a7af97e154698492fa5037e6d1df39c941aa1d911976709d2c8342253e33?s=96&d=mm&r=g\",\"contentUrl\":\"https:\\\/\\\/secure.gravatar.com\\\/avatar\\\/5c12a7af97e154698492fa5037e6d1df39c941aa1d911976709d2c8342253e33?s=96&d=mm&r=g\",\"caption\":\"Marc M\u00e9zard\"},\"url\":\"https:\\\/\\\/www.rivistaeco.com\\\/en\\\/author\\\/mmezard\\\/\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"Artificial Intelligence Supporting Science - Rivista Eco","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"http:\/\/www.rivistaeco.com\/en\/2024\/10\/06\/artificial-intelligence-supporting-science\/","og_locale":"en_US","og_type":"article","og_title":"Artificial Intelligence Supporting Science - Rivista Eco","og_description":"If everything can be explained by statistical correlations, what is the role today of the scientific method that has guided the development of our knowledge […]","og_url":"http:\/\/www.rivistaeco.com\/en\/2024\/10\/06\/artificial-intelligence-supporting-science\/","og_site_name":"Rivista Eco","article_published_time":"2024-10-06T15:19:14+00:00","author":"Marc M\u00e9zard","twitter_card":"summary_large_image","twitter_misc":{"Written by":"Marc M\u00e9zard","Est. reading time":"10 minutes"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"Article","@id":"http:\/\/www.rivistaeco.com\/en\/2024\/10\/06\/artificial-intelligence-supporting-science\/#article","isPartOf":{"@id":"http:\/\/www.rivistaeco.com\/en\/2024\/10\/06\/artificial-intelligence-supporting-science\/"},"author":{"name":"Marc M\u00e9zard","@id":"https:\/\/www.rivistaeco.com\/#\/schema\/person\/3e6fc92658259155b69a6065831f8e9e"},"headline":"Artificial Intelligence Supporting Science","datePublished":"2024-10-06T15:19:14+00:00","mainEntityOfPage":{"@id":"http:\/\/www.rivistaeco.com\/en\/2024\/10\/06\/artificial-intelligence-supporting-science\/"},"wordCount":1736,"articleSection":["Non categorizzato"],"inLanguage":"en-US"},{"@type":"WebPage","@id":"http:\/\/www.rivistaeco.com\/en\/2024\/10\/06\/artificial-intelligence-supporting-science\/","url":"http:\/\/www.rivistaeco.com\/en\/2024\/10\/06\/artificial-intelligence-supporting-science\/","name":"Artificial Intelligence Supporting Science - Rivista Eco","isPartOf":{"@id":"https:\/\/www.rivistaeco.com\/#website"},"datePublished":"2024-10-06T15:19:14+00:00","author":{"@id":"https:\/\/www.rivistaeco.com\/#\/schema\/person\/3e6fc92658259155b69a6065831f8e9e"},"breadcrumb":{"@id":"http:\/\/www.rivistaeco.com\/en\/2024\/10\/06\/artificial-intelligence-supporting-science\/#breadcrumb"},"inLanguage":"en-US","potentialAction":[{"@type":"ReadAction","target":["http:\/\/www.rivistaeco.com\/en\/2024\/10\/06\/artificial-intelligence-supporting-science\/"]}]},{"@type":"BreadcrumbList","@id":"http:\/\/www.rivistaeco.com\/en\/2024\/10\/06\/artificial-intelligence-supporting-science\/#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"http:\/\/www.rivistaeco.com\/en\/"},{"@type":"ListItem","position":2,"name":"Artificial Intelligence Supporting Science"}]},{"@type":"WebSite","@id":"https:\/\/www.rivistaeco.com\/#website","url":"https:\/\/www.rivistaeco.com\/","name":"Rivista Eco","description":"","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/www.rivistaeco.com\/?s={search_term_string}"},"query-input":{"@type":"PropertyValueSpecification","valueRequired":true,"valueName":"search_term_string"}}],"inLanguage":"en-US"},{"@type":"Person","@id":"https:\/\/www.rivistaeco.com\/#\/schema\/person\/3e6fc92658259155b69a6065831f8e9e","name":"Marc M\u00e9zard","image":{"@type":"ImageObject","inLanguage":"en-US","@id":"https:\/\/secure.gravatar.com\/avatar\/5c12a7af97e154698492fa5037e6d1df39c941aa1d911976709d2c8342253e33?s=96&d=mm&r=g9af4c0f16d62cde21b270763cad3dba3","url":"https:\/\/secure.gravatar.com\/avatar\/5c12a7af97e154698492fa5037e6d1df39c941aa1d911976709d2c8342253e33?s=96&d=mm&r=g","contentUrl":"https:\/\/secure.gravatar.com\/avatar\/5c12a7af97e154698492fa5037e6d1df39c941aa1d911976709d2c8342253e33?s=96&d=mm&r=g","caption":"Marc M\u00e9zard"},"url":"https:\/\/www.rivistaeco.com\/en\/author\/mmezard\/"}]}},"jetpack_featured_media_url":"","jetpack_sharing_enabled":true,"_links":{"self":[{"href":"https:\/\/www.rivistaeco.com\/en\/wp-json\/wp\/v2\/posts\/4547","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.rivistaeco.com\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.rivistaeco.com\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.rivistaeco.com\/en\/wp-json\/wp\/v2\/users\/7973"}],"replies":[{"embeddable":true,"href":"https:\/\/www.rivistaeco.com\/en\/wp-json\/wp\/v2\/comments?post=4547"}],"version-history":[{"count":1,"href":"https:\/\/www.rivistaeco.com\/en\/wp-json\/wp\/v2\/posts\/4547\/revisions"}],"predecessor-version":[{"id":4548,"href":"https:\/\/www.rivistaeco.com\/en\/wp-json\/wp\/v2\/posts\/4547\/revisions\/4548"}],"wp:attachment":[{"href":"https:\/\/www.rivistaeco.com\/en\/wp-json\/wp\/v2\/media?parent=4547"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.rivistaeco.com\/en\/wp-json\/wp\/v2\/categories?post=4547"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.rivistaeco.com\/en\/wp-json\/wp\/v2\/tags?post=4547"},{"taxonomy":"author","embeddable":true,"href":"https:\/\/www.rivistaeco.com\/en\/wp-json\/wp\/v2\/coauthors?post=4547"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}