Artificial intelligence, the reality behind the runaway – Pledge Times

Francis Bach is a research director at Inria, a national research institute in digital sciences and technologies, responsible for the Sierra team for statistical learning, one of the facets of artificial intelligence. He is a member of the Academy of Sciences. A pioneer and world-renowned specialist in machine learning, he is particularly interested in its algorithmic and theoretical aspects.

Everyone has been talking about artificial intelligence (AI) for the past ten years, in the scientific, economic and personal spheres. This renewal comes, like the previous waves of enthusiasm, with its promises (the autonomous car in ten years), its fears (will my job be replaced by a machine?), Even its fantasies (robots will they take control?).

Beyond this over-communication and all these expectations, AI is already among us in a very visible and impressive way, when we dictate a message to our phone, when we use an automatic translator, when our camera takes care of the Focus on faces or when the best chess and go players get beaten by machines.

AI is also present in a more hidden, even insidious, way when it is used for marketing campaigns, to follow us and offer us products to buy while we are browsing the Internet or to offer us personalized content on the various search engines, news article aggregators or social networks.

The fantasies of hostile robot takeover or transhumanism were quickly evacuated. Fears of threats to jobs are legitimate, but economic history teaches us that for general purpose technologies, like the steam engine, electricity, the computer or the Internet, and now the artificial intelligence, the effects on employment are multiple and difficult to predict, because, if some jobs disappear, new ones appear.

What about big promises, such as autonomous cars for all or personalized medicine? Why are the revolutionary applications promised by AI taking a long time to arrive? Mainly because current knowledge in computer science and mathematics is not sufficient. Research efforts are therefore crucial to achieve responsible and efficient AI.

Lets first dive into the factors behind the AI revival. These recent advances are largely the result of the systematic use of so-called machine learning algorithms, which use masses of data where many examples of the desired final behavior are available ( we speak of labeled data). Learning is generally supervised: the machine learns from examples, which themselves are provided by humans. For example, for the recognition of objects in images, it was necessary to provide by hand a description of the content of millions of images, and for machine translation, millions of sentences translated by humans. Learning like this from big data requires significant computing power.

By caricaturing, current AI technologies can only behave intelligently or almost in situations already observed previously and often enough. For object recognition in images or machine translation, existing tagged data or created for the occasion is sufficient, in non-critical contexts where errors made by AI are not of consequence. What about for future applications?

Take the example of the autonomous car. As the generalization capacities beyond what has been observed are currently limited, it is necessary to have proposed to the learning algorithms observations coming from landscapes (countryside, city or suburb), from different seasons (snow, rain, heatwave), different times of the day, different countries, or even the behavior styles of different road users (differentiated observation of the Highway Code), etc. To each of the previous criteria corresponds the need to multiply the number of data by a large factor (say arbitrarily 10, to fix minds). Because of this multiplicative effect, it will be tedious and expensive to acquire and label so much data.

AI is already with us in a very visible and impressive way, when we dictate a message to our phone, when our camera takes care of the focus on the faces or when the best chess and Go players are done beat by machines.

This is why, in most AI applications, new algorithms are needed to reduce the amount of tagged data, and use the available masses of data without supervision or with weak supervision (like a baby observes his environment) to understand how the world works or, more precisely, to create a representation of the world that allows learning with less labeled data.

The data collection necessary for learning algorithms, like any collection of personal data, poses difficult privacy issues. Do I have to give my explicit consent for my location and internet browsing data or the content of my emails to be used by large technology companies to serve me new advertisements? Is it really dangerous to share in a controlled, protected and anonymous way my recent physical contacts to participate in the fight against a pandemic? If I authorize the use of my medical data for epidemiological studies, will I benefit from potential new treatments? These questions concern all digital data processing and go beyond AI, but their upstream consideration by AI researchers is necessary for the widespread and responsible use of personalized algorithms.

The massive use of data poses another problem, of an ethical nature: that of biases in the data used and the difficulty of arriving at fair predictions. Take the example of an algorithm seeking to offer job offers from CVs: if it has learned from databases in the past, it will only offer jobs in the world of employment to men. computer science. There was a bias in the data, which the algorithm reproduced. In a medical application, if the panel of patients included in the initial study is not balanced, the diagnosis will not be as accurate for patients with under-represented characteristics. These biases are found at multiple levels, in particular social or ethnic. Taking them into account is an important current issue for machine learning, both in the development of new so-called fair algorithms and in their practical implementation.

The use of data processing in critical applications, such as transport or medicine, has long required the establishment of guarantees of proper functioning of the software used. In the context of learning algorithms, it is necessary to go beyond the computer bugs common to all software, which current certification technologies in principle make it possible to avoid. Indeed, learning methods are above all statistical methods for which the results contain margins of error (as for surveys). Quantifying this uncertainty is essential in critical systems, the great difficulties being that the final error rates which will be admitted by the users will be very low, and that the estimation of these uncertainties is for the moment out of reach for the users. latest algorithms that have enabled recent advances, such as deep neural networks.

Take the example of the algorithm seeking to offer job offers. If he learned from databases in the past, he will only offer computer jobs to men.

Beyond the natural variability of the data, the algorithms must also be robust to malicious attacks (for example ensuring that the photo of a person is not enough to unlock their phone by facial recognition or avoid being able to change a few pixels. chosen from an image in order to deceive the machine in its prediction). Here too, the development of robust and reliable methods is a particularly active field of research.

Processing massive amounts of data requires high computing power, which has an impact on the environment, notably through the construction of the necessary computers (energy expended, use of rare materials, transport and recycling), and the consumption of electricity. required to run and cool these computers.

Like all digital technologies, AI will replace existing activities that are costly for the environment. We naturally think of teleworking made possible by videoconferencing tools which, beyond health considerations, reduces daily transport, with, as with any technology, sometimes unexpected rebound effects (for teleworking, a move to a more distant home. workplace). Beyond the precise measurement of the impact of AI on the environmental level, it can only be positive if a more frugal AI is born: this requires a research effort that the machine learning community just starting to take into account.

Thus, for AI to benefit everyone, many scientific barriers will have to be lifted. It is difficult and probably illusory to want to make predictions on the performance of algorithms in ten years. On the other hand, it is now certain that a formidable research effort is necessary, for which well-funded academic research is essential to meet technical challenges, but also to inform citizens about the strengths and weaknesses of artificial intelligence.

Here is the original post:

Artificial intelligence, the reality behind the runaway - Pledge Times