ETIS Tandem Talk: Paul Gay & Guillaume Renton

Tandem Talk by Paul Gay (UPPA) and Guillaume Renton (ETIS) on Sustainability and Machine Learning.

Paul Gay | Interactions between sustainibility and machine learning research

AI is a controversial topic and different visions of sustainibility co-exist in the IT communities. In this talk, I will describe the state of the art from the GreenIT community to assess environmental impact of IT projects. Although the main ideas, such as life cycle analysis, description of embodied impacts and indirect effects have been explored for 10-20 years, we are only beginning to see them applied to AI systems and projects.

In a second part, I will present two of my current machine learning applications to sustainibility topics. The first one is the use of active learning to detect and classify controversial topic on renewable energies in social networks. The second one is to exploit the technique of early exit, an interesting tool where the amount of compute depends on the data, and which find applications in edge-cloud settings.

Guillaume Renton | Reducing computation costs without jeopardizing precision of Entity Alignment tasks

In recent work, we took interest in the computational cost of one of our entity alignment model, HybEA. The idea was to provide an efficiency analysis on top of a performance analysis, which is the main source of comparison between AI models. The efficiency analysis was conducted by using fvcore in order to compute the number of GFLOPS required to train the model. This has led to surprising results, showing that the initial embedding sizes of the models were oversized. This allowed us to greatly reduce the computational cost of our model with a very small loss of accuracy.