The main bottleneck for using deep neural networks in location-aided millimeter wave beam alignment procedures is the need for large datasets to tune their large set of trainable parameters. This letter proposes to use the transfer learning technique in order to reduce the dataset size requirements in deep-learning based beam selection. Information transfer can be done from one environment to another, or from one antenna configuration to another, which we refer to as domain and task adaptation, respectively. Numerical evaluations show a significant gain in using transfer learning in both domain and task adaptation scenarios, especially with limited datasets.
Paper
Full text
Deep Transfer Learning for Location-Aware Millimeter Wave Beam Selection
Semantic Scholar · Engineering · 2021
Abstract
The main bottleneck for using deep neural networks in location-aided millimeter wave beam alignment procedures is the need for large datasets to tune their large set of trainable parameters. This letter proposes to use the transfer learning technique in order to reduce the dataset size requirements in deep-learning based beam selection. Information transfer can be done from one environment to another, or from one antenna configuration to another, which we refer to as domain and task adaptation, respectively. Numerical evaluations show a significant gain in using transfer learning in both domain and task adaptation scenarios, especially with limited datasets.