Adaptation to Unknown Domains by Manipulating Activations of Neural Networks

Abstract

This dissertation presents a comprehensive study on enhancing the adaptation of neural networks to unknown domains by manipulating activations in computer vision tasks. The research addresses the critical challenge of distribution shifts between training and testing data, which significantly impacts the performance of deep neural networks in real-world scenarios. The dissertation introduces three novel methodologies designed to improve domain generalization and test-time adaptation. The first methodology dynamically generates plausible, novel, and diverse styles that are injected into the training data, introducing new feature statistics that were absent in the original data. This approach enhances the model's ability to generalize across various domains. The second methodology leverages lightweight yet informative proxies of training data by synthesizing condensed data and injecting the test data's style into it, creating test-stylized labeled data for virtually supervised test-time training. This method effectively utilizes the knowledge from the training data while minimizing memory overhead and privacy concerns. The third methodology, Memory-Based Batch Normalization (MemBN), reliably computes normalization statistics by aggregating previously stored in-batch statistics and combining them with source batch statistics. This computation for normalization statistics ensures precise feature distribution approximation, making the model more robust and efficient, particularly for on-device AI with memory constraints. The dissertation demonstrates the effectiveness of these methodologies through extensive experiments on various benchmarks, showing significant improvements in domain generalization and test-time adaptation research area. The proposed approach not only achieves high performance but also offers practical solutions for enhancing the adaptation ability of models in real-world environments. Overall, the contributions of this dissertation represent a significant advancement in enhancing the adaptation of neural networks, ensuring that models can effectively handle diverse testing environments encountered in real-world scenarios. By addressing the challenges of distribution shifts and leveraging innovative techniques, this research provides a solid foundation for future developments in robust and adaptable recognition models.