With the heterogeneity of modern computing architectures and the massive amount of data generated by edge devices, Machine Learning research is facing new challenges. Especially in time-critical or resource-constrained environments, targeted coordination of model and execution platform is indispensable. In particular, FPGAs (Field Programmable Gate Arrays) are well suited for IoT applications to save energy. At the same time, it is necessary to adapt new implementations of known ML algorithms to specific hardware requirements. To enable reliable application, resource requirements of several implementations for one algorithm need to be characterized. For example, even the well-known decision tree algorithm still offers novel research questions: Which implementation is best suited for which computing architecture and application requirements?

We investigate models of Machine Learning for distributed and strongly limited computer architectures and evaluate modern hardware regarding its suitability for Machine Learning. Learning procedures are examined and, if necessary, adapted to improve their executability on strongly resource-limited devices. Some of these adjustments go hand in hand with an approximation which must be theoretically justified and have error barriers. The results are used directly for algorithm curation, which includes describing limits of applicability as well as computational and energy requirements. Our current resource-aware ML research covers the following aspects:

Analysis of ML resource-requirements:

• • Resource-utilization labels including runtime, memory, and energy requirements of ML algorithms
  • Enable the certification process for resource-constrained and safety-critical ML applications

Investigating hard- and software for resource-constrained scenarios:

• • Resource-aware and distributed on-device learning
  • Decentralized data generation on edge devices
  • Resource-efficient implementations of anomaly detection algorithms
  • End-to-end learning from astroparticle image streams using deep learning on FPGAs
  • Image recognition-based tracking of logistic entities
  • C Code-generation for fast inference on IoT systems

Optimization of algorithms:

• • Resource-aware data acquisition through active class selection
  • Resource-aware hyperparameter-optimization for ML algorithms