Beschreibung
Neural Networks have revolutionized the artificial intelligence and machine learning field in recent years, enabling human and even super-human performance on several challenging tasks in a plethora of different applications. Unfortunately, these networks have dozens of millions of parameters and need billions of complex floating-point operations, which does not fit the requirements of rising Internet-of-Things (IoT) end nodes. In this work, these challenges are tackled on three levels: Efficient design and implementation of embedded hardware, the design of existing low-power microcontrollers and their underlying instruction set architecture, and full-custom hardware accelerator design. Meanwhile, we are investigating novel algorithmic approaches of extreme quantization of neural networks, and analyze their performance and energy efficiency trade-off.