Core Concepts
DNN models trained with fixed ETF classifiers improve transfer performance by minimizing class covariances, enhancing cluster separability, and focusing on essential features for class separation.
Abstract
DNN models trained with fixed ETF classifiers show significant improvement in transfer learning across domains by implicitly minimizing class covariances. The approach enhances cluster separability and focuses on essential features for better performance on out-of-domain datasets. By enforcing negligible within-class variability, the models achieve superior transfer performance compared to traditional methods.
The study explores the equivalence between Neural Collapse (NC) phenomenon and linear random features for classification robustness and generalization. Utilizing Random Matrix Theory results, the research establishes that linear random features exhibit minimal class covariance, leading to enhanced transfer performance.
Transfer learning experiments using ResNet50 and ResNet101 pretrained models demonstrate the effectiveness of DNNs trained with fixed ETF classifiers. The results highlight superior performance in out-of-domain scenarios, showcasing up to a 19% gain compared to baseline methods like Switchable Whitening (SW).
The practical implications of training DNN models with fixed ETF classifiers are evident in improved transferability across diverse data distributions. By implicitly minimizing class covariances during pretraining, the models focus on relevant features for class separation, reducing dependency on domain-specific variations.
Stats
Our approach outperforms baseline methods by up to 22% on fine-grained image classification datasets.
Methods explicitly whitening covariance during training show up to a 19% lower performance compared to our approach.
The model excels at adapting to different data distributions with gains of up to 19% over SW method.
In out-of-domain scenarios, our approach consistently outperforms baseline methods across all datasets and architectures.
The fixed ETF model shows a significant reduction in covariance compared to trainable and SW models after fine-tuning.
Quotes
"In contrast to numerous methods that explicitly whiten features covariances during training using dedicated loss functions, our approach focuses on enforcing negligible within-class variability throughout training."
"Our work bridges the application of fixed ETF classifiers with the use of Random Projection (RP) classifiers."
"The results presented highlight the effectiveness of utilizing DNN models trained with fixed ETF classifiers for transfer learning tasks."
"Our study extends the potential of fixed ETF classifiers by showcasing their effectiveness in cross-domain transfer tasks."
"Our research offers a perspective of the role played by fixed ETF classifiers in feature transformation and transfer learning."