Map Activation Quantization for Image Classification with Limited Data

Deep learning applications in image classification often rely on extensive training data. However, in specific scenarios, obtaining a substantial amount of training data poses significant challenges. Training models with limited data often leads to overfitting and poor generalization performance. To address this issue, we propose a novel method based on map activation quantization for image classification. Our method quantizes spatial activations of latent representations, modeling the relationship between discrete representations and categories. This effectively captures variations in object properties, providing more reliable classification information and mitigating overfitting in scenarios with limited data. Furthermore, our approach seamlessly integrates with any convolutional neural network model without necessitating alterations to existing architectures or methodologies. Empirical evaluations on natural and medical image datasets demonstrate the superiority of our method in image classification under limited data scenarios, establishing a new technological benchmark. We intend to publicly release the relevant source code.