What is AI?

WHAT IS ARTIFICIAL INTELLIGENCE? Villani mission on artificial intelligence March 2018

Defining artificial intelligence (AI) is no Grounded on logic easy matter. The field is so broad that it Some researchers such as John cannot be limited to a specific area of McCarthy, also one great figure in the research. AI is more of an ambition: it AI field, argue that rational logic is a seeks to understand how human more relevant “standard” of cognition works by creating cognitive intelligence measurement than human processes that emulate those of human capabilities. This approach to AI takes beings. advantage of the tools offered by mathematical logic to formalize the AI is at the crossroad of multiple fields: complex tasks to be accomplished by computer science, mathematics (logics, artificial intelligence machines. The optimisation, analysis, probabilities, main issue this approach has to deal linear algebra…), cognitive science… with is the formalization of the tasks. For These core scientific disciplines also example, we can ask ourselves which need to be mixed with the specific apprehensible rules of reasoning allow knowledge of the fields they are us to distinguish almost immediately a applied to, and each algorithm in AI is picture of a cat from that of a dog supported by a mix of techniques: among the hundreds of possible races? semantic analysis, symbolic computing, Moreover, how to take into account the machine learning, exploratory analysis, fact that the image could be blurred? deep learning and neural networks… Besides, in our daily decisions, uncertainty is constantly present and Defining AI managed by other rules than logic: Imitate human capabilities habit, emotions and intuitions play an essential role in these mechanisms. Marvin Lee Minsky, who is considered What is the goal of AI? as one of the founding fathers of AI, defines it as follows: “the science of One could wonder why we even bother making machines do things that would to build smart machines. The primary require intelligence if done by men. It goal would be that they make our lives requires high-level mental processes easier by taking care of complex and such as: perceptual learning, memory repetitive daily tasks: find our best and critical thinking.” itineraries to reach our friend’s house, In other words, artificial intelligence is select the most relevant content the science of building computer according to our tastes, translate programs that aim to perform tasks that texts… That is what experts refer to would require some intelligence if they when they talk about “weak AI”. were done by human beings. Therefore, These programs are only able to no human activity seems to be out of perform one single and delimited task, reach: moving from one place to hence the “weak” qualification. another, learning, reasoning, Algorithms building our favorite socializing, creativity etc. Nevertheless, playlists, finding the most relevant we are still far from creating a machine search results or identifying our friends that would be able to match or on social media are not able to do outperform human capabilities in all anything more than this. Yet, this does fields. not make these algorithms any less useful, but it could lead to some artificial stupidity: if one AI technique is

trained to distinguish cats versus dogs, by a physician are transcribed into code if provided a human face picture it will to automatically produce diagnoses. still answer either dog or cat without As an a priori approach, these flinching. techniques require to split into fixed To the other end sits what is called classes the objects processed by the “strong AI”, able to match and even algorithms according to some outperform human capabilities in all predefined features. As such, it is very fields. We are still very far from building challenging to capture the complexity such a machine, and any unconvinced of the real world and establish fixed reader just has to engage in a short categories. discussion with the most advanced Machine learning chatbot to be fully convinced. The fact that the prospect of a “strong AI” Alongside symbolic AI, machine machine is highly unlikely anytime soon learning techniques were developed in does not prevent people, experts and order to model cognitive process regular citizens alike, from passionately directly from previous experience. debating its destructive or positive Every machine learning technique has impact. It is no accident that the sci-fi two steps: the first one is the learning culture, which has been central to the phase that uses the input data (e.g. development of AI, is full of this kind of pictures of cats and dog for the argument. Further, previous classification task) to find the technological advances also raised parameters that best fit the task at similar fears, and it is not surprising that hand. The second step takes the we observe the same phenomenon with learned parameters as input and AI. Nevertheless, policymakers and performs the task accordingly. This is regulators are already urged to the inference phase. consider the fundamental questions brought about by the development of Among machine learning approach, we AI; especially considering diverse can identify a few broad classes of projections regarding its potential problems. dramatic effect on employment. That’s a tricky matter for regulators; little Supervised learning accustomed to such speculative topics. In supervised learning, one learns a set How to build AI? of rules in order to perform a specific task according to a given set of Symbolic AI: grounded on logic examples for which we are given the expected result. For example, we could Symbolic AI refers to a class of be given a million of pictures of cats and approach to AI that aims to reproduce dogs identified as such. The learning the reasoning mechanisms of humans. phase progressively changes the These mechanisms, put in logical form, parameters of the model in order to are produced by a priori modeling of make the error as small as possible on knowledge and rules of logic to perform the known (and unknown) examples. In the task at hand. One of the the inference phase, the algorithm distinguished illustrations of symbolic performs the task on unknown data— AI technique is the expert systems e.g. it classifies dogs and cats on algorithm, capable of producing previously unseen pictures. reasoning from known facts and rules. For instance, diagnostic rules defined

Unsupervised learning In unsupervised learning, we only get Deep learning unlabeled data with no additional Deep learning is a subfield of machine knowledge. In the learning phase, the learning that has come to a new era goal is to find the underlying structure since 2006. These models are built from of the data such as categories. The basics components called “neurons” inference phase is the same as the that are organized into successive supervised one. The unsupervised layers. These neurons are linked by approach is especially useful to find connections whose weight is adjusted in behavioral profiles from the recorded the learning phase. Every neuron maps activity without being given beforehand its input to its output with a transfer any prior knowledge about these function which is schematically inspired profiles. by the brain, but it is well known that These approaches are at the core of this is a crude analogy. current research projects since their success would enable AI techniques to After the learning phase, the neural rely less on the availability of labeled network is theoretically able to split the data, which is very expensive to input data into a hierarchy of features produce. representing multiple abstraction layers. For example in the case of face Reinforcement learning recognition, the first layer identifies the Reinforcement learning is different from elementary patterns such as lines, supervised and unsupervised learning edges, corners, the next ones find large because it does not rely on any pre- patterns (lips, forehead, eyes…), and existing set of examples as input. the description is progressively refined Instead, leaning happens directly (size of the nose…) until the description through interactions with the is complete. environment: the machine performs an action, the state of the environment Neural networks are the go-to changes according to this action, and approach in signal processing (sound, the machine gets a reward depending images, language, video etc) since they on the result. are able to extract complex spatial and temporal structures. The reward enables the machine to discover the best actions to perform Two kinds of models are especially well according to the result it aims to get. A known: convolutional networks recent and famous example of (especially in computer vision) and reinforcement learning is AlphaGo: recurrent networks for time series and here, the action performed is a move on language processing applications, in the board, the environment is the board which the notion of layer is a little with the stones on it, and the reward is fuzzier. the outcome of a game, either win or loss.

Applications Let’s finish with some concrete Then, a neural network goes process examples. the texts to find statistical patterns. Image and video recognition From these patterns, it builds equivalences between sequences of Interpreting an image—recognizing a words, which allow it to translate an person or an object and its surrounding unknown text. environment—is a relatively easy task However, translation softwares have for a human being. Every day, our brain fewer documents translated for some process effortlessly complex visual language pairs. For instance, there are information: a family picture, a car, a more documents translated from landscape. However, it is a very French into Spanish than from Danish to challenging task for a computer. Romanian. That’s why the quality of Yet the stakes are high because the translations varies a lot from one development of autonomous cars (for language to another. the perception of its surrounding Content recommendation environment), the automating labeling of images, the improvement of To recommend content to their users, identification systems, the detection of online marketplace or streaming pathologies from medical imaging platforms, use AI systems which devices are all dependent on advances operate according to a different in image and video recognition. approach from those of the learning How do social media platforms techniques exposed above. Here, input recognize faces on pictures? data are composed of all the past choices made by users. This dataset is They use convolutional neural networks. used to create fake user profiles and According to the same principle of product categories, here “fake” means supervised learning for the classification that they represent average user of the pictures of dogs and cats behaviours. Because of the wide variety described above, the system learns to of users and products, it is impossible to distinguish the faces of user’s friends for infer consumption behaviours whom he has labeled images. When a categories directly from their photo is uploaded to the platform, the consumption choices. Each user can system only has to categorize the new then be analyzed from these fake faces present by matching them with categories, which makes it possible to the labelled faces from its database. compute a “fake proximity” between Translation users more reliably than simply counting the products they ordered. In order to build a translation software, Afterwards, the system can recommend we start by building a large database of to a specific user a set of products likely texts translated by human translators to match his/her preferences because that serve as models. These translations they have been selected according to often come from books, official similar fake users' preferences. documents produced by international organizations (United Nations, European Commission) and authoritative websites. We talk about millions of texts…