TY - GEN
T1 - Counterfactual Contrastive Learning for Fine Grained Image Classification
AU - Yin, Chenke
AU - Wang, Jia
AU - Zhang, Haichao
AU - Feng, Kaiyue
AU - Shi, Lin
AU - Ma, Qianyi
N1 - Publisher Copyright:
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2024.
PY - 2024
Y1 - 2024
N2 - In the realm of fine-grained image classification, discerning subtle distinctions between closely related categories remains a challenge. However, these approaches typically fall short in addressing the deeper causal relationships that underlie the visible features, leading to potential biases and limited generalizability. This paper presents a fine-grained causal contrastive network (FCCN), a novel architecture that integrates causal inference with contrastive learning to address the intricacies of fine-grained visual classification. Unlike traditional approaches that predominantly rely on feature correlations, FCCN deeply studies the causal relationship between salient features and labels, which greatly enhances the model’s ability to discriminate fine-grained images. A key innovation of FCCN is the introduction of a backdoor adjustment technique for feature decoupling, effectively minimizing the impact of irrelevant context feature and purifying the feature space for more precise classification. We validate our model on CUB-200-2011, Stanford Cars, and WM-811K datasets. Both accuracy and robustness are significantly improved, which demonstrate notable improvements in accuracy and robustness.
AB - In the realm of fine-grained image classification, discerning subtle distinctions between closely related categories remains a challenge. However, these approaches typically fall short in addressing the deeper causal relationships that underlie the visible features, leading to potential biases and limited generalizability. This paper presents a fine-grained causal contrastive network (FCCN), a novel architecture that integrates causal inference with contrastive learning to address the intricacies of fine-grained visual classification. Unlike traditional approaches that predominantly rely on feature correlations, FCCN deeply studies the causal relationship between salient features and labels, which greatly enhances the model’s ability to discriminate fine-grained images. A key innovation of FCCN is the introduction of a backdoor adjustment technique for feature decoupling, effectively minimizing the impact of irrelevant context feature and purifying the feature space for more precise classification. We validate our model on CUB-200-2011, Stanford Cars, and WM-811K datasets. Both accuracy and robustness are significantly improved, which demonstrate notable improvements in accuracy and robustness.
KW - Causal inference
KW - Explainable AI
KW - Fine-grained visual classification
UR - http://www.scopus.com/inward/record.url?scp=85205320082&partnerID=8YFLogxK
U2 - 10.1007/978-3-031-72341-4_12
DO - 10.1007/978-3-031-72341-4_12
M3 - Conference Proceeding
AN - SCOPUS:85205320082
SN - 9783031723407
T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
SP - 169
EP - 183
BT - Artificial Neural Networks and Machine Learning – ICANN 2024 - 33rd International Conference on Artificial Neural Networks, Proceedings
A2 - Wand, Michael
A2 - Schmidhuber, Jürgen
A2 - Wand, Michael
A2 - Malinovská, Kristína
A2 - Schmidhuber, Jürgen
A2 - Tetko, Igor V.
A2 - Tetko, Igor V.
PB - Springer Science and Business Media Deutschland GmbH
T2 - 33rd International Conference on Artificial Neural Networks, ICANN 2024
Y2 - 17 September 2024 through 20 September 2024
ER -