What is computational reproducibility?

International audienceComputational modelling is the process by which phenomena found in complex systems are expressed algorithmically. The creation of such simulations is useful because it allows us to test whether our understanding is sophisticated enough to create credible working models of the phenomena we are studying. In neuroscience and cognitive science especially, computational modelling comprises more than just capturing a single phenomenon, it also implements a theory. It gives scientists a method of allowing their ideas to be executed, i.e., for emergent properties to appear when they are implemented and run (McClelland, 2009). In this context, a model is said to be replicable if experiments within it can be carried out successfully using the original codebase, with the implicit assumption that such a codebase is available. However, for models to be evaluated it is mandatory to ensure they are reproducible (Topalidou, Leblois, Boraud, & Rougier, 2015). That is, that they can be recreated based on their specification — the details deemed important enough to be included in the accompanying article (Hinsen, 2015). Ideally, this should be possible without contacting the authors for advice, and critically, without referring to the original code (Cooper & Guest, 2014). If the specification is sufficient to successfully recreate the codebase from scratch, then the model is said to be reproducible. This adds further credence to both the model and its overarching theoretical framework. If not, and the model cannot be recreated, then even if the experiments can be carried out successfully within the original codebase, the model is not reproducible (Crook, Davison, & Plesser, 2013). How to share computational research