Informatics Colloquium 25.2. Energy-efficient neural networks for embedded systems
Informatics Colloquium 25.2. 2019, 14:00 lecture hall D2
Ing. Vojtěch Mrázek, Ph.D., FIT VUT
Energy-efficient neural networks for embedded systems
Abstract: Artificial neural networks are optimized for high-performance computer
systems. However, the inference path of the NNs is often executed in small
embedded systems such as special ASIC or FPGA accelerators. Since these systems
are typically battery-powered, the energy consumption becomes crucial. In this
context, this talk deals with three topics. (i) Overview and challenges of
hardware NN accelerators. (ii) Error resiliency of neural networks. (iii)
Approximations of NN inference path for applications such as low power image
classifiers.