Context Adaptive Edge Inference
Bachelor’s Thesis / Master’s Thesis / Student Research Project
Abstract
DNN models that are executed are subject to a number of conditions and characteristics that contextualize them and their execution with regards to both hardware and software. For example, achievable classification accuracy might depend on the available hardware, i.e., it depends on hardware context. Similarly, achievable accuracy might be influenced by varying levels of input data quality, or latency might be influenced by a varying number of layers, both possible cases of software context dependency. When it comes to scheduling on distributed, heterogeneous hardware that can change its configuration over time, wether it be a single DNN model or several, informational exchange about hardware and software contexts becomes vital with regards to finding optimal schedules. Furthermore, in resource constrained, real-time critical edge environments (in sensor proximity), communication has to be both quick and energy efficient, e.g. motivating the development of a portable context information exchange method. Device contexts like cache size and memory hierarchy can impact possible software optimizations like matrix tiling.
References
- Paper: “Using the Abstract Computer Architecture Description Language to Model AI Hardware Accelerators”
- Paper: “AdaMEC: Towards a context-adaptive and dynamically combinable DNN deployment framework for mobile edge computing”
- IREE
Requirements
- Python
- Linux and Git
- C / Assembly (recommended)
- Understanding of deep neural networks
- Understanding of computer architectures
- Successfully atteded the lecture “Grundlagen der Rechnerarchitektur” (recommended)
- Successfully atteded the lecture “Efficient Machine Learning in Hardware” (recommended)
- Successfully attended th lecture “Modellierung und Analyse von Eingebetteten Systemen” (recommended)