Navigating the Depths of Adaptive Embedded Intelligence: Ensembling, Reconfiguring and Editing Models

Activity: Talk or presentation › Invited talk › Science to science

Description

Deep models often lack the flexibility needed for dynamic adaptation, such as responding to fluctuating resource availability or input data transformations. Adaptation is particularly difficult on resource-constrained devices. In this scenario, linear functions are crucial due to their compatibility with hardware parallelism and accelerator capabilities. Our research tackles this by discovering and leveraging linear mode connectivity (LMC) of deep neural networks. LMC, a phenomenon where trained networks are connected via a linear path with non-increasing loss, is pivotal for ensembling, reconfiguring, and editing models. This talk delves into the origin, scope, and implications of LMC, demonstrating how it enables lightweight and adaptive embedded intelligence.
Period14 Dec 2023
Event titleAVL TechTrend
Event typeSeminar
LocationGraz, AustriaShow on map
Degree of RecognitionLocal

Related content