Riccardo Zecchina, Director of the Department of Computing Sciences at Bocconi University, delves into the intricate connections between artificial intelligence, computer science, statistical physics, and computational neuroscience.<\/span><\/p>\n This column explores the technology that is transforming the world. As AI evolves daily, simply defining its scope is a complex matter,. Let us then begin with its history, to clarify how it differs from human intelligence, and what the future might entail.<\/span><\/i><\/p>\n In this column, we will discuss artificial intelligence, explaining in simple terms how it is <\/span>transforming our world<\/b>. We will cover topics ranging from the ways in which intelligent machines operate, the differences with human and animal thinking, right up to their impact on society.<\/span><\/p>\n Let’s start by clarifying what we mean by “artificial intelligence” today, the key moments in its development, and what we might expect from it in the near future; for making predictions that go beyond a few years is risky, even for experts in the field.\u00a0<\/span><\/p>\n So, what does \u201cartificial intelligence\u201d mean today?<\/span><\/p>\n Today, “artificial intelligence” does not refer to the thinking machines seen in science fiction movies. Today’s AI involves computers that learn directly from examples, with little to no human guidance.\u00a0<\/span><\/p>\n For instance, instead of writing a set of rules to recognize the ears or whiskers of a cat in a photo\u2014a method that would lead to many misclassifications, such as confusing cats with dogs\u2014we now have <\/span>computers that learn to identify cats by analyzing millions of images<\/b>, and that often outperform humans.<\/span><\/p>\n Today’s artificial intelligence is based on data learning. This is how we are able to translate languages, convert speech to text, and create new content, including articles, programming codes, music, and images, just to name a few. These tools are accelerating development in all sectors in which data is crucial, from basic science to societal technologies.<\/span><\/p>\n Today’s artificial intelligence is thus synonymous with “machine learning.”<\/b> When referring to machines capable of thinking like humans, we speak of “general artificial intelligence,” a field that remains beyond our reach despite advances in machine learning.<\/span><\/p>\n The history of modern artificial intelligence is a fascinating one that has developed in a series of succeeding waves, corresponding to the increase in computing power and available data. In simpler terms, simple models attempting to mimic the operation of individual brain neurons were introduced in the 1950s. These early ideas allowed computers to perform basic tasks, such as distinguishing male from female faces, however their predictive capabilities remained were very limited.<\/span><\/p>\n The 1980s saw a shift towards more complex systems, such as networks composed of multiple neurons (artificial neural networks). For example, the so-called convolutional networks significantly improved image recognition yet failed to become the standard for all applications, often remaining a second choice.<\/span><\/p>\n Between 2005 and 2015, new computing platforms (originally developed for other purposes, such as video games), coupled with the explosion of data available on the web, allowed for the exploring of neural networks of unprecedented size. Such networks, initially met with skepticism, definitively demonstrated their potential in 2012, when they surpassed every other algorithm in terms of image recognition. This success not only secured definitive support from the scientific community, but also attracted the attention of American tech giants.<\/span><\/p>\n Between 2015 to 2020, new types of networks were introduced that could not only recognize and classify, but also generate new informational content. These are the <\/span>“generative models,” with large language models like ChatGPT being perhaps the most well-known example<\/b>. This technology has since been employed in other fields, from biomedical research to software and drug development.<\/span><\/p>\n We are now in the era of multimodality, where different types of data, such as text, images, and sensory information, are combined in order to further improve the capabilities of generative AIs.<\/span><\/p>\n Currently, generative neural networks are of a staggering size; for instance, Gpt-4 has about two trillion connections, and its training requires enormous computational resources, far beyond the reach of most universities and companies. Concurrently, more manageable open-source generative models are being developed, which could one day become valuable tools for research and practical applications.<\/span><\/p>\n Unlike other scientific fields in which theory guides practical development, in artificial intelligence, expensive experimentation currently precedes theory. This forces academic institutions and companies to constantly chase after a field that is endlessly and rapidly evolving.<\/span><\/p>\n But how does machine learning work?\u00a0<\/span><\/p>\n Imagine wanting to teach a machine how to understand and interpret the world: before even wondering how it might “think,” it is imperative to ensure it understands the stimuli it receives and learns from them. This may be biological for us and animals, but it has always been a great challenge for a traditional computer.\u00a0<\/span><\/p>\n Here’s how it works for machines, specifically in artificial neural networks:<\/span><\/p>\n This is the essence of machine learning: an endless cycle of trials and errors <\/b>that enables the machine to “understand” data and make predictions, oftentimes even very reliable ones.\u00a0<\/span><\/p>\n The quality of the predictions heavily depends on the quantity of data available and on the type of goal set. For instance, artificial neural networks perform better than humans when recognizing objects within images. This is even done on large scales, as seen in Massachusetts where they are used for mammograms.<\/span><\/p>\n The use of language generative systems such as ChatGPT is particularly topical. Think of artificial neural networks as very diligent students, who have spent countless hours reading every single thing that has ever been written: books, articles, blog posts, the equivalent of hundreds of millions of books. They learn to understand the use of words, a sentence\u2019s structure and the way in which concepts are expressed. They learn the ways in which words entwine, understanding both grammar and context. Once they have acquired such ability, they are able to produce texts, images and sounds that resemble those created by humans. The \u201cmagic\u201d behind the process is once again machine learning, albeit a particular form of it, known as \u201cself-supervised\u201d learning: machines take a piece of text, hide some words, and modify their connective paths in order to guess the words that had been previously hidden. By doing this, the network develops the ability to complete sentences and predict the upcoming words: a system that has the power to generate coherent and plausible texts is created, one that works even if simply given a beginning, or a prompt.<\/span><\/p>\n Will these systems ever become as intelligent as humans<\/b>? Probably not, as they lack many of the characteristics that make up human or animal intelligence, such as knowledge of the physical world and the ability to imagine scenarios. Through models based solely on language processing, we cannot simulate very effectively how we and animals learn outside of the use of language. At least for now.\u00a0<\/span><\/p>\n The fundamental difference between human learning and that of language models is that we do not read millions of books (a thousand would already be quite an achievement); we learn much more from observing and interacting with the real world. Despite this, it is fascinating to wonder whether the expressive power encapsulated in the human language might be enough to achieve an understanding of the world.\u00a0<\/span><\/p>\n The debate is open: is intelligence based on perceived reality, or can language suffice? Reality is rich and complex, and planning actions without abstract models of the world seems difficult. AIs excel over humans in complex strategy games, translating and composing texts, and even acing professional exams. Yet, they still struggle with tasks that appear trivial to us, such as learning how to drive.\u00a0<\/span><\/p>\n In the future, multimodality, meaning training with videos and images in addition to text, will further Ai\u2019s improvements by rendering the system more versatile. However, one question remains: will this approach lead us towards more advanced forms of artificial intelligence, ones that are capable of planning and performing complex actions? That is the challenge that awaits us.<\/span><\/p>\n","protected":false},"excerpt":{"rendered":" A Few Words on Artificial Intelligence Riccardo Zecchina, Director of the Department of Computing Sciences at Bocconi University, delves into the intricate connections between artificial […]<\/p>\n","protected":false},"author":5739,"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":[64],"class_list":["post-868","post","type-post","status-publish","format-standard","hentry","category-non-categorizzato"],"acf":[],"yoast_head":"\nWhat AI is Today<\/b><\/h3>\n
A Brief History of Artificial Intelligence<\/b><\/h3>\n
Machine Learning<\/b><\/h3>\n
\n
Human Intelligence<\/b><\/h3>\n