A Theoretical Explanation for Perplexing Behaviors of Backpropagation-based Visualizations

Authors: Weili Nie, Yang Zhang, Ankit Patel

ICML 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments are provided that support the theoretical analysis. We do extensive experiments to support our theory and further reveal more detailed properties of these backpropagation-based visualizations
Researcher Affiliation Academia 1Department of Electrical and Computer Engineering, Rice University, Houston, USA. 2Department of Computer Science, Rice University, Houston, USA. 3Department of Neuroscience, Baylor College of Medicine, Houston, USA. Correspondence to: Weili Nie <wn8@rice.edu>, Ankit B. Patel <abp4@rice.edu>.
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code Yes Code is available at https://github.com/weilinie/BackpropVis
Open Datasets Yes Unless stated otherwise, the input is the image tabby from the Image Net dataset (Deng et al., 2009) with size 224 224 3.
Dataset Splits No The paper mentions using ImageNet for experiments and refers to a test set for calculating l2 distance statistics (“10K images from the Image Net test set”), but it does not specify explicit training, validation, and test dataset splits or their sizes/percentages needed for reproduction.
Hardware Specification No The paper mentions “computational limitations” for certain experiments (e.g., downsampling input image size to 64x64x3), but it does not specify any concrete hardware details such as GPU models, CPU types, or memory used for the experiments.
Software Dependencies No The paper does not provide specific software dependencies or version numbers for libraries, frameworks, or solvers used in the experiments (e.g., Python, PyTorch, TensorFlow versions).
Experiment Setup Yes For a random network, their weights are all sampled from the truncated Gaussians with a zero-mean and standard deviation 0.1. In the three-layer CNN, the filter size is 7 7 3, the number of filters is N = 256, and the stride is 2. In the three-layer FCN, the hidden layer size is set to Nh = 4096. By default, the backpropagation-based visualizations are calculated with respect to the maximum class logit.