Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Smoothed Differentiation Efficiently Mitigates Shattered Gradients in Explanations
Authors: Adrian Hill, Neal McKee, Johannes Maeß, Stefan Bluecher, Klaus-Robert Müller
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate Smooth Diff s excellent speed and performance in a number of experiments and benchmarks. Thus, Smooth Diff greatly enhances the usability (quality and speed) of Smooth Grad a popular workhorse of XAI. |
| Researcher Affiliation | Academia | 1BIFOLD Berlin Institute for the Foundations of Learning and Data, Berlin, Germany 2Machine Learning Group, TU Berlin, Berlin, Germany 3Bernstein Center for Computational Neuroscience, Berlin, Germany 4TU Berlin, Berlin, Germany 5Department of Artificial Intelligence, Korea University, Seoul, Korea 6Max Planck Institut für Informatik, Saarbrücken, Germany |
| Pseudocode | Yes | Pseudocode implementation for both Re LUs and max pooling layers are given in Appendix I. |
| Open Source Code | Yes | We provide the complete source code to reproduce our experiments at https://github.com/ adrhill/smoothdiff-experiments, including Smooth Diff reference implementations in Julia [40] and Py Torch. |
| Open Datasets | Yes | We evaluate Smooth Diff and SG on a pre4trained VGG419 model [14] and the Image Net dataset [33]. |
| Dataset Splits | No | The paper uses pre-trained models and various sets of images for evaluation (e.g., "a batch of 128 randomly drawn Image Net images", "batches of 128 Image Net images", "a dataset of 256 input images") but does not specify traditional training/test/validation splits for model training or for the evaluation of the explanation methods themselves. |
| Hardware Specification | Yes | All experiments were run on an NVIDIA A100 80GB GPU and AMD EPYC 9124 CPU. |
| Software Dependencies | No | We use the Julia-XAI ecosystem [34, 35], Flux.jl deep learning framework [36, 37] and implement VEJPs using Chain Rules.jl [38] and Zygote.jl [39]. |
| Experiment Setup | Yes | A standard deviation of 𝜎= 0.5 is used for the sampling distribution 𝒩𝑔 in both Smooth Diff and SG. (...) Table 1: Pixel4flipping results (...) Smooth Diff 𝜎= 0.5, 𝑛= 4 |