Debunking Grad-ECLIP: A Comprehensive Study on Its Incorrectness and Fundamental Principles for Model Interpretation
Not on the map. Why?
Abstract (arXiv)
Grad-ECLIP is published at ICML 2024 and represents a new Transformer interpretation technical route (intermediate features-based). First, this paper demonstrates that the intermediate features-based technical route is not a novel one. Based on the existing attention-based route, we have developed Attention-ECLIP, which is completely equivalent to Grad-ECLIP but with simpler computation. Both through formal derivation and experimental validation, we prove that the intermediate feature-based route represented by Grad-ECLIP is actually an equivalent variant of the attention-based route. Next, this paper demonstrates that the Grad-ECLIP method is flawed. The model interpretation results obtained by Grad-ECLIP are not those of the original model, and the interpretation results are misaligned with the model's performance. We analyze the causes of Grad-ECLIP's flaws and propose, or rather, explicitly emphasize two fundamental principles that model interpretation should adhere to in order to avoid similar errors.
Key facts. Cited by 2 papers in the map (data through 7 Oct 2026). Builds on 31 of its 44 references that are in the map. First posted on arXiv 13 May 2026.
Builds on
31 of 44 references.
References · 31
- 2017Attention Is All You Needcited by 59,932
- 2020An Image is Worth 16x16 Words: Transformers for Image Recognition at Scalecited by 29,915
- 2014Microsoft COCO: Common Objects in Contextcited by 17,498
- 2021Swin Transformer: Hierarchical Vision Transformer using Shifted Windowscited by 12,082
- 2014ImageNet Large Scale Visual Recognition Challengecited by 8,441
- 2016Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localizationcited by 6,690
- 2016"Why Should I Trust You?": Explaining the Predictions of Any Classifiercited by 5,040
- 2022BLIP: Bootstrapping Language-Image Pre-training for Unified Vision-Language Understanding and Generationcited by 4,310
- 2015Learning Deep Features for Discriminative Localizationcited by 2,430
- 2021Learning to Prompt for Vision-Language Modelscited by 2,269
- 2019Do ImageNet Classifiers Generalize to ImageNet?cited by 1,507
- 2019VisualBERT: A Simple and Performant Baseline for Vision and Languagecited by 1,291
19 more
- 2021Conceptual 12M: Pushing Web-Scale Image-Text Pre-Training To Recognize Long-Tail Visual Conceptscited by 1,209
- 2019Are Sixteen Heads Really Better than One?cited by 803
- 2020Quantifying Attention Flow in Transformerscited by 759
- 2021CLIP4Clip: An Empirical Study of CLIP for End to End Video Clip Retrievalcited by 645
- 2020Transformer Interpretability Beyond Attention Visualizationcited by 620
- 2021RegionCLIP: Region-based Language-Image Pretrainingcited by 612
- 2022GroupViT: Semantic Segmentation Emerges from Text Supervisioncited by 529
- 2021Generic Attention-model Explainability for Interpreting Bi-Modal and Encoder-Decoder Transformerscited by 316
- 2023Efficient and Effective Text Encoding for Chinese LLaMA and Alpacacited by 302
- 2020UNIMO: Towards Unified-Modal Understanding and Generation via Cross-Modal Contrastive Learningcited by 239
- 2021CoTr: Efficiently Bridging CNN and Transformer for 3D Medical Image Segmentationcited by 227
- 2016Layer-wise Relevance Propagation for Neural Networks with Local Renormalization Layerscited by 197
- 2019Iterative Answer Prediction with Pointer-Augmented Multimodal Transformers for TextVQAcited by 107
- 2020Developing Real-time Streaming Transformer Transducer for Speech Recognition on Large-scale Datasetcited by 94
- 2021Multi-Compound Transformer for Accurate Biomedical Image Segmentationcited by 42
- 2017Interpretability via Model Extractioncited by 40
- 2022ViT-CX: Causal Explanation of Vision Transformerscited by 27
- 2016TreeView: Peeking into Deep Neural Networks Via Feature-Space Partitioningcited by 9
- 2023Deep Integrated Explanationscited by 9
Built on by
Citing papers per year: 2026: 2.
2 papers.