Code for Overinterpretation paper Overinterpretation reveals image classification model pathologies

Overview

Overinterpretation

This repository contains the code for the paper:

Overinterpretation reveals image classification model pathologies
Authors: Brandon Carter, Siddhartha Jain, Jonas Mueller, David Gifford

Introduction

Image classifiers are typically scored on their test set accuracy, but high accuracy can mask a subtle type of model failure. We find that high scoring convolutional neural networks (CNNs) on popular benchmarks exhibit troubling pathologies that allow them to display high accuracy even in the absence of semantically salient features. When a model provides a high-confidence decision without salient supporting input features, we say the classifier has overinterpreted its input, finding too much class-evidence in patterns that appear nonsensical to humans. Here, we demonstrate that neural networks trained on CIFAR-10 and ImageNet suffer from overinterpretation, and we find models on CIFAR-10 make confident predictions even when 95% of input images are masked and humans cannot discern salient features in the remaining pixel-subsets. Although these patterns portend potential model fragility in real-world deployment, they are in fact valid statistical patterns of the benchmark that alone suffice to attain high test accuracy. Unlike adversarial examples, overinterpretation relies upon unmodified image pixels. We find ensembling and input dropout can each help mitigate overinterpretation.

Usage

Dependencies

Python 3.7
PyTorch v1.5.0
torchvision v0.5.0

Full requirements in requirements.txt.

Overview

The overinterpretation pipeline can be understood as:

  1. Train models on full images (train.py).
  2. Run backward selection for all training and test images (run_sis_on_cifar.py).
  3. Train new models on pixel-subsets of images and mask the remaining pixels (train.py).
  4. Evaluate new models and compare accuracy to original models.

The relevant scripts for running this pipeline are train.py and run_sis_on_cifar.py. Each script contains usage examples in the docstring. train.py supports training models on full image data as well as pixel-subsets only (specified via command line arguments, usage examples in docstring).

Note that for CIFAR-10, when training models on pixel-subsets only, we keep 5% of pixels and mask the remaining 95% with zeros.

Citation

If you use our methods or code, please cite:

@inproceedings{overinterpretation,
  title={Overinterpretation reveals image classification model pathologies},
  author={Carter, Brandon and Jain, Siddhartha and Mueller, Jonas W and Gifford, David},
  journal={Advances in Neural Information Processing Systems},
  volume={34},
  year={2021}
}
Owner
Gifford Lab, MIT CSAIL
Gifford Lab, MIT CSAIL
The source code for Generating Training Data with Language Models: Towards Zero-Shot Language Understanding.

SuperGen The source code for Generating Training Data with Language Models: Towards Zero-Shot Language Understanding. Requirements Before running, you

Yu Meng 38 Dec 12, 2022
Kaggle DSTL Satellite Imagery Feature Detection

Kaggle DSTL Satellite Imagery Feature Detection

Konstantin Lopuhin 206 Oct 29, 2022
GndNet: Fast ground plane estimation and point cloud segmentation for autonomous vehicles using deep neural networks.

GndNet: Fast Ground plane Estimation and Point Cloud Segmentation for Autonomous Vehicles. Authors: Anshul Paigwar, Ozgur Erkent, David Sierra Gonzale

Anshul Paigwar 114 Dec 29, 2022
Official repository for the paper "Going Beyond Linear Transformers with Recurrent Fast Weight Programmers"

Recurrent Fast Weight Programmers This is the official repository containing the code we used to produce the experimental results reported in the pape

IDSIA 36 Nov 15, 2022
Imagededup - 😎 Finding duplicate images made easy

imagededup is a python package that simplifies the task of finding exact and near duplicates in an image collection.

idealo 4.3k Jan 07, 2023
Source code for the Paper: CombOptNet: Fit the Right NP-Hard Problem by Learning Integer Programming Constraints}

CombOptNet: Fit the Right NP-Hard Problem by Learning Integer Programming Constraints Installation Run pipenv install (at your own risk with --skip-lo

Autonomous Learning Group 65 Dec 27, 2022
[NeurIPS 2021] Deceive D: Adaptive Pseudo Augmentation for GAN Training with Limited Data

Near-Duplicate Video Retrieval with Deep Metric Learning This repository contains the Tensorflow implementation of the paper Near-Duplicate Video Retr

Liming Jiang 238 Nov 25, 2022
Simple renderer for use with MuJoCo (>=2.1.2) Python Bindings.

Viewer for MuJoCo in Python Interactive renderer to use with the official Python bindings for MuJoCo. Starting with version 2.1.2, MuJoCo comes with n

Rohan P. Singh 62 Dec 30, 2022
The Few-Shot Bot: Prompt-Based Learning for Dialogue Systems

Few-Shot Bot: Prompt-Based Learning for Dialogue Systems This repository includes the dataset, experiments results, and code for the paper: Few-Shot B

Andrea Madotto 103 Dec 28, 2022
Project page for End-to-end Recovery of Human Shape and Pose

End-to-end Recovery of Human Shape and Pose Angjoo Kanazawa, Michael J. Black, David W. Jacobs, Jitendra Malik CVPR 2018 Project Page Requirements Pyt

1.4k Dec 29, 2022
Detecting Human-Object Interactions with Object-Guided Cross-Modal Calibrated Semantics

[AAAI2022] Detecting Human-Object Interactions with Object-Guided Cross-Modal Calibrated Semantics Overall pipeline of OCN. Paper Link: [arXiv] [AAAI

13 Nov 21, 2022
Optimizaciones incrementales al problema N-Body con el fin de evaluar y comparar las prestaciones de los traductores de Python en el ámbito de HPC.

Python HPC Optimizaciones incrementales de N-Body (all-pairs) con el fin de evaluar y comparar las prestaciones de los traductores de Python en el ámb

Andrés Milla 12 Aug 04, 2022
Plugin adapted from Ultralytics to bring YOLOv5 into Napari

napari-yolov5 Plugin adapted from Ultralytics to bring YOLOv5 into Napari. Training and detection can be done using the GUI. Training dataset must be

2 May 05, 2022
Official source code to CVPR'20 paper, "When2com: Multi-Agent Perception via Communication Graph Grouping"

When2com: Multi-Agent Perception via Communication Graph Grouping This is the PyTorch implementation of our paper: When2com: Multi-Agent Perception vi

34 Nov 09, 2022
Official implementation of "Motif-based Graph Self-Supervised Learning forMolecular Property Prediction"

Motif-based Graph Self-Supervised Learning for Molecular Property Prediction Official Pytorch implementation of NeurIPS'21 paper "Motif-based Graph Se

zaixi 71 Dec 20, 2022
SAT: 2D Semantics Assisted Training for 3D Visual Grounding, ICCV 2021 (Oral)

SAT: 2D Semantics Assisted Training for 3D Visual Grounding SAT: 2D Semantics Assisted Training for 3D Visual Grounding by Zhengyuan Yang, Songyang Zh

Zhengyuan Yang 22 Nov 30, 2022
ColBERT: Contextualized Late Interaction over BERT (SIGIR'20)

Update: if you're looking for ColBERTv2 code, you can find it alongside a new simpler API, in the branch new_api. ColBERT ColBERT is a fast and accura

Stanford Future Data Systems 637 Jan 08, 2023
Turning SymPy expressions into JAX functions

sympy2jax Turn SymPy expressions into parametrized, differentiable, vectorizable, JAX functions. All SymPy floats become trainable input parameters. S

Miles Cranmer 38 Dec 11, 2022
Hough Transform and Hough Line Transform Using OpenCV

Hough transform is a feature extraction method for detecting simple shapes such as circles, lines, etc in an image. Hough Transform and Hough Line Transform is implemented in OpenCV with two methods;

Happy N. Monday 3 Feb 15, 2022
Consumer Fairness in Recommender Systems: Contextualizing Definitions and Mitigations

Consumer Fairness in Recommender Systems: Contextualizing Definitions and Mitigations This is the repository for the paper Consumer Fairness in Recomm

7 Nov 30, 2022