What are the major challenges currently facing artificial intelligence?
In my area of expertise, which is machine learning (ML), the three topics that I am currently passionate about, and that could potentially be considered as the great challenges in this field, are bias and fairness, weak signals, and learning on networks. But this is only a partial view of the challenges in AI, which is a very broad and mostly interdisciplinary field. AI is a set of tools, methods and technologies that enable a system to perform tasks in a quasi-autonomous way, and there are different ways of achieving this.
ML is about the machine learning from examples, training itself to perform tasks efficiently that it will later undertake. The great successes in this area are computer vision and automatic listening, used for applications in biometrics for example, and natural language processing. One of the questions that currently arises is how much confidence can be placed in ML tools, as deep learning requires very large volumes of data, which very often come from the web.
Unlike datasets previously collected by researchers, web data is not acquired in a “controlled” way. The vast quantity of this data can sometimes lead to the methodological questions that should be asked to exploit the information it contains being ignored. For example, training a facial recognition model directly from web data can lead to bias, in the sense that the model would not recognise all types of faces with the same efficiency. In this case, the bias may stem from a lack of representativeness in the faces used.
If, for example, the data corresponds mainly to Caucasian faces, the system developed may recognise Caucasian faces more easily than other types of faces. However, the disparities in performance may also be due to the intrinsic difficulty of the prediction problem and/or the limitations of current ML techniques: it is well known, for example, that deep learning does perform as will in the recognition of the faces of new-borns as it does for adult faces. However, there is currently no clear theoretical insight into the link between the structure of the deep neural network used and the performance of the model for a given task.
You say “currently”. Does that mean that these biases could one day be removed, or their effect could diminish?
There are different types of bias. They can be relative to the data, there are the so-called “selection” biases, linked to the lack of representativeness, “omission” biases, due to errors through endogeneity, etc. Biases are also inherent in the choice of the neural network model, the ML method, a choice that is inevitably restricted to the state of the art and limited by current technology. In the future, we may use other, more efficient, less computationally intensive representations of information, which can be more easily deployed, and which may reduce or eliminate these biases, but for the moment they exist!
What role does the quality of the datasets used for training play in these biases?
It is very important. As I said, given the volume of data required, it is often sourced from the web and therefore not acquired in a sufficiently controlled way to ensure representativeness. But there is also the fact that this data can be ‘contaminated’, in a malicious way. This is currently an issue for the computer vision solutions that will be used in autonomous vehicles. The vehicle can be deceived by manipulating the input information. It is possible to modify the pixel image of, say, a traffic sign so that the human eye sees no difference, but the neural network ‘sees’ something other than the traffic sign.
ML is based on a frequentist principle and the question of the representativeness of the data in the learning phase is a major issue. Using autonomous driving as an example, we now see many vehicles on the Saclay plateau, equipped with sensors to store as much experience as possible. That being said, it is difficult to say how long it will be before we have seen enough situations to be able to deploy a sufficiently intelligent and reliable system in this field, enabling us to deal with all future situations.
There are certainly applications for which the data available today allows ML to be implemented in a satisfactory manner. This is the case, for example, for handwriting recognition, for which neural networks are perfectly developed. For other problems, in addition to experimental data, generative models will also be used, producing artificial data that account for adverse situations but without claiming to be exhaustive. This is the case for ML applications in cybersecurity, in an attempt to automatically detect malicious intrusions into a network for example.
Generally speaking, there are many problems for which the data available is too sparse to implement ML in a simple way. This is often the case in anomaly detection, particularly for predictive maintenance of complex systems. In some cases, the hybridisation of ML and symbolic techniques in AI could provide solutions. These avenues are being explored in the civil and military aviation sectors, as well as in medical imaging. In addition to their effectiveness, such approaches may also enable machines to make decisions that are easier to explain and interpret.
What is driving evolution in AI today?
The field of mathematics contributes a lot, especially in terms of efficient information representation and algorithms. But it is also technological progress that is driving AI forward. The mathematical concept of neural networks has been around for many decades. Recent technical advances, particularly in the field of memory, have made it possible to successfully implement deep neural network models. Similarly, distributed computing architectures and dedicated programming frameworks have made it possible to scale up learning on large volumes of data. What remains to be done is to design more frugal approaches, so as to reduce the carbon footprint of computations, which is a very topical issue!