Neural networks in the automotive sector commonly have to process varying number of objects per observation. Deep Set feature extractors have shown great success on problems in reinforcement learning with dynamic observations, achieving state-of-the-art performance on common tasks like highway driving. However, recent work has shown that neural networks suffer from a spectral bias, fitting to low frequencies of scalar features used in such representations, like velocities and distances. We introduce a novel set feature extractor combining learned Fourier features with the Deep Set architecture. The proposed architecture reduces this spectral bias allowing more sample-efficient training and better performance by learning a more detailed representation. Extensive experiments are conducted on three different environments, including two novel environments featuring a large number of objects in challenging random scenarios. Our method outperforms state-of-the-art approaches and exceeds the performance of existing Deep Set architectures on these challenging new tasks.
Learned Fourier Bases for Deep Set Feature Extractors in Automotive Reinforcement Learning
2023-09-24
1157779 byte
Conference paper
Electronic Resource
English
Feature Extraction and Image Registration of Color Images using Fourier Bases
British Library Conference Proceedings | 2005
|