Resource-Aware Machine Learning

Resource-aware machine learning has two main motivations. On the one hand, the internet of things, embedded systems, edge {AI}, and federated learning demand machine learning to manage computation with less resources, i.e., runtime, memory, communication, and energy. On the other hand, learning large models need to become more aware of resources, because they consume too much. Regarding the climate change, saving resource consumption has become an urgent need. Both motivations lead to the same scientific subject, namely the design and implementation of machine learning algorithms that are optimized to get along with less resources than a straight-forward version. Where embedded systems always dealt with various computing architectures, the larger models and finally the large language models rely on efficient chips with parallel processing. In any case, the implementation on a certain hardware needs to be taken into account. Given the huge environmental impact of computing, the choice of an implemented model should now be based on how low its resource consumption is. Hence, it is important to measure, test, and report model features such that users can easily compare the implemented models and choose the one with a minimal footprint.