Learning explanations that are hard to vary

Giambattista Parascandolo, Alexander Neitz, Antonio Orvieto, Luigi Gresele, Bernhard Schölkopf
9/1/2020
cs.LGstat.ML

Official code: pytorch/ignite (4,795 stars, BSD-3-Clause, PyTorch, last commit 2026-10-20). 1 more implementation below.

Abstract

In this paper, we investigate the principle that `good explanations are hard to vary' in the context of deep learning. We show that averaging gradients across examples -- akin to a logical OR of patterns -- can favor memorization and `patchwork' solutions that sew together different strategies, instead of identifying invariances. To inspect this, we first formalize a notion of consistency for minima of the loss surface, which measures to what extent a minimum appears only when examples are pooled. We then propose and experimentally validate a simple alternative algorithm based on a logical AND, that focuses on invariances and prevents memorization in a set of real-world tasks. Finally, using a synthetic dataset with a clear distinction between invariant and spurious mechanisms, we dissect learning signals and compare this approach to well-established regularizers.

View on arXivView PDF

Code Implementations(2)

pytorch/igniteOfficial100%
PyTorchBSD-3-Clause

High-level library to help with training and evaluating neural networks in PyTorch flexibly and transparently.

4,795746Apex, ShellNov 23, 20175 days agoBSD-3-Clause
closemberdeep-learninghacktoberfestmachine-learningmetrics+3 more
gibipara92/learning-explanations-hard-to-varyOfficial100%

Code to implement the AND-mask and geometric mean to do gradient based optimization, from the paper "Learning explanations that are hard to vary"

415Python, Jupyter NotebookOct 16, 20205 years ago

Similar Papers

The Pitfalls of Memorization: When Memorization Hurts Generalization

82% match1 repo412024

Learning with Explanation Constraints

82% match2023

Per-Example Gradient Regularization Improves Learning Signals from Noisy Data

82% match2023

Cross-Entropy Is All You Need To Invert the Data Generating Process

82% match2024

Good-looking but Lacking Faithfulness: Understanding Local Explanation Methods through Trend-based Testing

82% match1 repo42023

HyperInvariances: Amortizing Invariance Learning

82% match2022

Cite this paper

@article{parascandolo2020learning,
  title  = {Learning explanations that are hard to vary},
  author = {Giambattista Parascandolo and Alexander Neitz and Antonio Orvieto and Luigi Gresele and Bernhard Schölkopf},
  year   = {2020},
  eprint = {2009.00329},
  archivePrefix = {arXiv},
  url    = {https://arxiv.org/abs/2009.00329}
}

Add a badge to your README

Shows whether official code is linked for this paper.

Code of Paper badge
[![Code of Paper](https://codeofpaper.com/badge/2009.00329.svg)](https://codeofpaper.com/paper/2009.00329)

Discussion