The experimental approach in digital sciences: between simulation, validation and interdisciplinarity

A mathematical or computer model can explain a phenomenon and predict its behaviour, but it is only through experience that it can be validated or invalidated. This approach is based on a fundamental principle of the scientific method: falsifiability, theorised by Karl Popper. According to this principle, a hypothesis must be refutable to be considered scientific, and is only valid as long as no experiment or observation has contradicted it. When an anomaly is detected, a new explanatory framework must be proposed, sometimes leading to scientific revolutions.

The Mnemosyne project team is interested in modelling the brain and the various interactions between neuronal structures to produce cognitive behaviour. "The models we design are compared with the results of animal experiments carried out on rodents and primates in collaboration with certified professionals. This supervision enables us to test our hypotheses by manipulating learning conditions and to compare them with analyses of the animals' brain responses," explains Nicolas Rougier, the team's research director. Each new experiment allows us to refine the models, reject certain scenarios and propose new ones.

Scientific experimentation can take different forms depending on the field of study and the objectives pursued. Some research requires a totally controlled environment, where every parameter can be adjusted to test precise hypotheses. Others, on the other hand, take place in natural conditions, as close as possible to reality, even if this means accepting a degree of unpredictability. Finally, the digital sciences offer a third type of experiment, where the phenomena being studied are entirely simulated in computer environments.

Laboratory experiments make it possible to rigorously control experimental variables and ensure reproducible conditions. The Pleiade project team, which focuses on bioinformatics, works with biologists, for whom only experimental evidence is accepted. We work with teams in biology and health, fields where only experimentation can establish proof," explains David Sherman, team leader. While we require mathematical demonstrations to validate the correctness of our methods, we have to recognise that, for our partners, this does not constitute proof in itself. This is why we invest a great deal of effort in ensuring that our methods provide explanations that can generate hypotheses that can be tested in multi-omics and reproduced.

Conversely, some research requires experiments in natural conditions, where control over the environment is more limited but the results can be more relevant. This is the case with the studies carried out by the Flowers AI & CogSci project team, which is exploring learning by studying the cognitive development of children. "By testing our models in situations that children encounter on a daily basis, such as visiting schools, we are able to gain a better understanding of natural learning mechanisms and draw lessons for the development of artificial intelligence systems inspired by living organisms, which in turn could be useful for developing educational technologies," explains Pierre-Yves Oudeyer, head of the team.

Finally, digital experimentation is another type of approach, specific to the digital sciences. In some cases, the field of study is itself a computer environment. The PlaFRIM platform provides an experimental space for testing high-performance computing algorithms and refining simulation models. Similarly, the Canari project team is developing algorithms in cryptography and number theory. In this field, experimentation consists of optimising these tools to guarantee both performance and security. These experiments, although carried out in virtual environments, are essential for validating methods used on a large scale in fields as varied as cybersecurity, physical modelling and artificial intelligence.

Reproducibility is one of the fundamental pillars of the scientific approach: an experiment must be repeatable under similar conditions and lead to the same results in order to be considered valid. However, this principle comes up against a number of difficulties depending on the discipline. The diversity of study areas, the complexity of the phenomena observed, technical constraints and the biases of experimental cohorts sometimes make reproducibility difficult to achieve in practice. This is particularly the case with the Bivwac and Potioc project teams, which are carrying out human-machine interaction experiments on human cohorts. These cohorts can be biased in their composition: age, education or the environment in which the subjects live, all of which influence their behaviour and abilities.

Faced with these challenges, the teams at the Inria Centre at the University of Bordeaux are paying particular attention to implementing rigorous practices that guarantee the reproducibility of experiments. 

Faced with these challenges, the research teams at the Inria Centre at the University of Bordeaux are putting in place rigorous practices to encourage open science and the reproducibility of experiments. But scientific experimentation is not limited to setting up protocols and analysing results: it also raises ethical and transparency issues (governed by the Coerle at Inria) that directly influence the way in which research is conducted and disseminated.

In a world where fake news and scientific misinformation are spreading rapidly, the experimental approach plays a vital role in the sciences as a means of comparing reality and theory. Thanks to its infrastructures, its varied experimental facilities and its wide range of technical and scientific skills, the Inria Centre at the University of Bordeaux is a key player in this scientific process, in a number of application areas. This expertise, combining modelling, simulation and experimentation, encourages an interdisciplinary approach and multiple collaborations in the Nouvelle-Aquitaine region, as well as nationally and internationally, guaranteeing the exchange of knowledge and the creation of new insights.