Denoising Diffusion Path: Attribution Noise Reduction with An Auxiliary Diffusion Model

Authors: Yiming Lei, Zilong Li, Junping Zhang, Hongming Shan

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experimental results demonstrate that DDPath can significantly reduce noise in the attributions resulting in clearer explanations and achieves better quantitative results than traditional path-based methods. 5 Experiments 5.1 Experimental Setup Datasets. Following previous studies [10, 11, 12], we evaluate the effectiveness of DDPath on the validation set of Image Net-1k [30] that contains 50, 000 images of 1, 000 classes. Furthermore, we conducted a pointing game experiment on MS COCO validation set [31].
Researcher Affiliation Academia Yiming Lei1, Zilong Li1, Junping Zhang1, Hongming Shan2 1 Shanghai Key Laboratory of Intelligent Information Processing, School of Computer Science, Fudan University 2 Institute of Science and Technology for Brain-Inspired Intelligence & MOE Key Laboratory of Computational Neuroscience and Brain-Inspired Intelligence & MOE Frontiers Center for Brain Science, Fudan University
Pseudocode Yes Algorithm 1 Algorithm of DDPath-IG. Require: Target image x and its label y, the initial noisy baseline x randomly sampled from a Gaussian; target model F( ), diffusion trained classifier hϕ, the diffusion model Eθ, total number of step T. Return: Attribution for the target image x: A = 1 T PT 1 t=0 gt.
Open Source Code No The data employed in this study are the publicly available, and the underlying source code will be released later.
Open Datasets Yes Datasets. Following previous studies [10, 11, 12], we evaluate the effectiveness of DDPath on the validation set of Image Net-1k [30] that contains 50, 000 images of 1, 000 classes. Furthermore, we conducted a pointing game experiment on MS COCO validation set [31].
Dataset Splits Yes Datasets. Following previous studies [10, 11, 12], we evaluate the effectiveness of DDPath on the validation set of Image Net-1k [30] that contains 50, 000 images of 1, 000 classes. Furthermore, we conducted a pointing game experiment on MS COCO validation set [31].
Hardware Specification Yes All the experiments are implemented by Py Torch [34] and conducted on an NVIDIA A100 GPU.
Software Dependencies No All the experiments are implemented by Py Torch [34] and conducted on an NVIDIA A100 GPU. (Only PyTorch is mentioned, without a specific version number).
Experiment Setup Yes The number of sampling steps for DDPath methods is 250. The above experiments involving DDPath used ρ = 1 t T , κ = t T by default, which maintains a stably decreased weight for sampling mean. Here, we reverse such scaling as ρ = t T , κ = 1 t T to enable a smaller mean and larger variance at the initial steps.