Speaker
Description
For machine learning applications on edge devices, inference speed and hardware resource usage are often limiting factors. These challenges can be mitigated by using model compression techniques such as quantization and pruning. However, these approaches introduce additional hyperparameters that require optimization. Hyperparameter optimization has been widely used to design models with the best prediction accuracy. In this study, we explore the possibility of using multi-objective hyperparameter optimization to include latency and hardware resource usage to the optimization process in addition to the prediction accuracy. In our study, we introduce a custom loss function to the optimization process. The function incorporates distinct terms for validation loss, latency, and resource usage, with adjustable coefficients for the latency and resource usage to control their relative importance. The term for resource usage is made up of an exponential penalty function, enabling the available resources to be used efficiently. Our findings suggest that the custom loss function enables optimizing a model architecture to specific hardware specifications.
| Would you like to be considered for an oral presentation? | No |
|---|
