The ICML (International Conference on Machine Learning) is one of the most important and renowned conferences in machine learning and artificial intelligence. This year, it took place in Vienna – and covered a wide range of topics, including algorithms, theoretical work, and concrete application examples.
The conference is a forum for discussing scientific work submitted by researchers from around the world. The best and most interesting papers are selected for presentation at the conference – this year, the work of HPI doctoral student Nicolas Alder was among them.
Nicolas presented his paper "Energy-Efficient Gaussian Processes Using Low-Precision Arithmetic" in Vienna. He is working on ideas for saving energy when using artificial intelligence. This research area is becoming increasingly important as the use of AI models grows. The HPI doctoral student took a closer look at the special method of Gaussian process regression in this context. This is a method that helps to make predictions in the field of artificial intelligence.
It can be imagined as a particularly clever guess based on past observations.
Nicolas Alder is investigating what happens when you use less accurate numbers. He also investigates how this lower accuracy affects the results of the Gaussian process regression and how the properties of the data, the type of implementation, the performance of the model and the energy consumption are related. His results show that under certain conditions, energy consumption can be reduced by up to 89.01% for almost all computational operations without affecting the performance of the model.
To put this in perspective, this means that even with less accurate figures, the accuracy of the predictions remains almost the same, while the energy consumption is significantly reduced. This could make artificial intelligence applications significantly more energy-efficient.
If you are interested in Nicolas Alder's work, here is the link to his paper:https://proceedings.mlr.press/v235/alder24a.html
Last change: 12/09/2024