Tools and Principles

One of the core values of iAIMS is open science research. Especially as AI is at risk of fuelling a new reproducibility crisis like we have seen in psychology (cfr. this Nature article), it is important to ensure reproducibility of our experiments. Just some examples of why AI research might not be reproducible:

• Different test sets are used by different studies
• Retraining a model with all the same configurations and resources probably results in a different model as many optimisation functions rely on a degree of randomness (e.g. Stochastic Gradient Descent)
• Data sets are often not open source, or not even with conditional access.

To maximise reproducibility, Open Science is a great leap forward! In this Wiki Page, we offer practical guidance.

Git

Git has become the bread and the butter of code version controlling and is widely used to publish code alongside a research paper.