The AI4Accelerators team of KIT participated in the AI Hero Hackathon 2023!! We joined the research field Energy, and the goal was to tune a Mask R-CNN model to identify thermal leaks from images of buildings in Karlsruhe. The caveat was to maintain accuracy while minimizing the energy needed to develop and train the model (sent to the GPUs of the HoreKa supercomputer facility) and also minimizing the energy for inference. As researchers we strived for the best accuracy (and our model was indeed the most accurate!) but at the price of an increased energy consumption. It was really interesting to be exposed to the actual energy that training large models require, and we came back with a long list of strategies to reduce it 🤓
