Sensor-Guided Adaptive Data Processing using Butterfly Coefficients from Hyper Networks Deep learning is gaining popularity in edge applications, but it comes with its own set of issues: On the one hand side, the unknown characteristics of an edge environment require more generalized models, which are larger and require more computational resources. On the other hand side, edge devices are more restrained in their computational power and resources and thus require smaller models. We introduce a novel approach where we use Hyper Networks to generate Butterfly Matrices for sensor-guided adaptive data processing. Butterfly Matrices are an efficient way of representing structured matrices using a butterfly-like shape for its coefficients. Hyper Networks are a novel type of neural network, which does not learn how to solve a given task, but yields a neural network as to solve that given task. In this work, we use Hyper Networks to generate Butterfly Matrices which are specifically crafted for the environment of the edge device, allowing for efficient data processing on the edge.
If you would like to attend the talk, please sign-up via office@csh.ac.at
Franz Papst (TU Graz, CSH) will give a talk on Friday January 13, 2023 at 3 PM in room 201 in a hybrid mode.