This meeting between CSH and IST Austria (Institute of Science and Technology Austria) will take place on Friday, July 1st at the CSH to discuss challenges and achievements of Deep Neural Networks and Green AI.
Abstract:
Three factors drive success of deep neural networks (DNNs): algorithmic innovation, growing data volumes, and the amount of available compute power. The latter two are responsible for the bigger-is-better race in deep learning and for a number of associated risks with this technology, in particular its environmental footprint and low data efficiency.
Sparsity promises to make DNN training and inference resource-aware and green by reducing model size and improving inference time. However, the performance of sparse models often differs from their dense versions. Sparsity may improve model performance, yet introduces unwanted biases when it comes to generalization and robustness.
Recent research results reveal striking differences in the behavior of sparsification methods and settings with respect to properties of the optimized models. The goal of this workshop is to improve our understanding of model sparsification by jointly reviewing recent empirical results and the metrics used to assess model performance. This workshop also promotes collaboration among Austrian researchers working towards green AI.
Organizer:
Confirmed participants:
Rahim Entezari (TU Graz, CSH)
Franz Papst (TU Graz, CSH)
Dan Alistarh (IST Austria)
Xiaoxi He (ETH Zürich)