• Corpus ID: 231918418

A Too-Good-to-be-True Prior to Reduce Shortcut Reliance

@article{Dagaev2021ATP,
  title={A Too-Good-to-be-True Prior to Reduce Shortcut Reliance},
  author={Nikolay I. Dagaev and Brett D. Roads and Xiaoliang Luo and Daniel N. Barry and Kaustubh R. Patil and Bradley C. Love},
  journal={ArXiv},
  year={2021},
  volume={abs/2102.06406}
}
Despite their impressive performance in object recognition and other tasks under standard testing conditions, deep networks often fail to generalize to out-of-distribution (o.o.d.) samples. One cause for this shortcoming is that modern architectures tend to rely on “shortcuts” – superficial features that correlate with categories without capturing deeper invariants that hold across contexts. Real-world concepts often possess a complex structure that can vary superficially across contexts, which… 
Methods for Estimating and Improving Robustness of Language Models
Despite their outstanding performance, large language models (LLMs) suffer notorious flaws related to their preference for simple, surface-level textual relations over full semantic complexity of the
Improved Worst-Group Robustness via Classifier Retraining on Independent Splits
TLDR
This work develops a method, called CRIS, that improves upon state-of-the-art methods, such as Group DRO, on standard datasets while relying on much fewer group labels and little additional hyperparameter tuning.
Last Layer Re-Training is Sufficient for Robustness to Spurious Correlations
TLDR
It is demonstrated that simple last layer retraining on large ImageNet-trained models can match or outperform state-of-the-art approaches on spurious correlation benchmarks, but with profoundly lower complexity and computational expenses.
ITSA: An Information-Theoretic Approach to Automatic Shortcut Avoidance and Domain Generalization in Stereo Matching Networks
TLDR
It is shown that using the proposed ITSA approach, state-of-the-art stereo matching networks that are trained purely on synthetic data can effectively generalize to challenging and previously unseen real data scenarios and outperform their finetuned counterparts for challenging out-ofdomain stereo datasets.
Encouraging Intra-Class Diversity Through a Reverse Contrastive Loss for Single-Source Domain Generalization
TLDR
A publicly available MNIST-based benchmark is designed to precisely measure the ability of an algorithm to find the "hidden" patterns and a partially reversed contrastive loss is proposed to encourage intra-class diversity and find less strongly correlated patterns.
Simple data balancing achieves competitive worst-group-accuracy
TLDR
The results show that these data balancing baselines achieve state-of-the-art-accuracy, while being faster to train and requiring no additional hyper-parameters.
Identifying and Benchmarking Natural Out-of-Context Prediction Problems
TLDR
This work introduces a framework unifying the literature on OOC performance measurement, and demonstrates how rich auxiliary information can be leveraged to identify candidate sets of OOC examples in existing datasets.
Focus on the Common Good: Group Distributional Robustness Follows
TLDR
Group-DRO is worse than ERM on four of the tasks shown in the table, including the in-domain (sub-population shift) evaluation on PovertyMap task, and CGD is the only algorithm that performs consistently well across all the tasks.
Encouraging Intra-Class Diversity Through a Reverse Contrastive Loss for Better Single-Source Domain Generalization
TLDR
This paper designs two simple MNIST-based SDG benchmarks and proposes a partially reversed contrastive loss to encourage intra-class diversity and less strongly correlated patterns, to deal with SDG-MP, and evaluates several state-of-the-art SDG algorithms through the authors' simple benchmark.
Competency Problems: On Finding and Removing Artifacts in Language Data
TLDR
This work argues that for complex language understanding tasks, all simple feature correlations are spurious, and formalizes this notion into a class of problems which are called competency problems, and gives a simple statistical test for dataset artifacts that is used to show more subtle biases.
...
...

References

SHOWING 1-10 OF 51 REFERENCES
Deep Residual Learning for Image Recognition
TLDR
This work presents a residual learning framework to ease the training of networks that are substantially deeper than those used previously, and provides comprehensive empirical evidence showing that these residual networks are easier to optimize, and can gain accuracy from considerably increased depth.
When Do Curricula Work?
TLDR
The experiments demonstrate that curriculum, but not anti-curriculum can indeed improve the performance either with limited training time budget or in existence of noisy data, suggesting that any benefit is entirely due to the dynamic training set size.
Learning from others' mistakes: Avoiding dataset biases without modeling them
TLDR
This work considers cases where the bias issues may not be explicitly identified, and shows a method for training models that learn to ignore these problematic correlations, based on the observation that models with limited capacity primarily learn to exploit biases in the dataset.
Learning to Model and Ignore Dataset Bias with Mixed Capacity Ensembles
TLDR
This paper proposes a method that can automatically detect and ignore dataset-specific patterns, which it hypothesize are likely to reflect dataset bias, and trains a lower capacity model in an ensemble with a higher capacity model.
Self-Challenging Improves Cross-Domain Generalization
TLDR
A simple training heuristic, Representation Self-Challenging (RSC), is introduced that significantly improves the generalization of CNN to the out-of-domain data and is presented theoretical properties and conditions of RSC for improving cross-domain generalization.
Learning from Failure: Training Debiased Classifier from Biased Classifier
TLDR
This work intentionally train the first network to be biased by repeatedly amplifying its ''prejudice'', and debias the training of the second network by focusing on samples that go against the prejudice of the biased network in (a).
What shapes feature representations? Exploring datasets, architectures, and training
TLDR
This work finds that when two features redundantly predict the label, the model preferentially represents one, and its preference reflects what was most linearly decodable from the untrained model.
The Pitfalls of Simplicity Bias in Neural Networks
TLDR
It is demonstrated that common approaches for improving generalization and robustness---ensembles and adversarial training---do not mitigate SB and its shortcomings, and a collection of piecewise-linear and image-based datasets that naturally incorporate a precise notion of simplicity and capture the subtleties of neural networks trained on real datasets are introduced.
Shortcut Learning in Deep Neural Networks
TLDR
A set of recommendations for model interpretation and benchmarking is developed, highlighting recent advances in machine learning to improve robustness and transferability from the lab to real-world applications.
...
...