Journal: Pattern Recognition Letters

Article Title: A too-good-to-be-true prior to reduce shortcut reliance ☆

doi: 10.1016/j.patrec.2022.12.010

Figure Lengend Snippet: The standard and too-good-to-be-true prior approaches to learning. (A) In the standard approach, a single high-capacity network (HCN) is trained and is susceptible to shortcuts, in this case relying on color as opposed to shape. Such a network will generalize well to i.i.d. test items but fail on o.o.d. test items (the last item for each class; shown in red). (B) In contrast, implementing the too-good-to-be true prior by pairing a low-capacity network (LCN) with an HCN leads to successful i.i.d. and o.o.d. generalization. Items that the LCN can master, which may contain shortcuts, are downweighted when the HCN is trained, which should reduce shortcut reliance and promote use of more complex and invariant features by the HCN. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Article Snippet: We trained an LCN (softmax regression; ) and an HCN (ResNet-56; ) to classify the colored MNIST dataset.

Techniques:

Journal: Pattern Recognition Letters

Article Title: A too-good-to-be-true prior to reduce shortcut reliance ☆

doi: 10.1016/j.patrec.2022.12.010

Figure Lengend Snippet: LCN and HCN mean accuracies (in %; averaged across 10 independent runs) on colored MNIST; standard deviations are in parentheses.

Article Snippet: We trained an LCN (softmax regression; ) and an HCN (ResNet-56; ) to classify the colored MNIST dataset.

Techniques:

Journal: Pattern Recognition Letters

Article Title: A too-good-to-be-true prior to reduce shortcut reliance ☆

doi: 10.1016/j.patrec.2022.12.010

Figure Lengend Snippet: LCN and HCN mean accuracies (in %; averaged across 10 independent runs) on stylized Tiny ImageNet; standard deviations are in parentheses.

Article Snippet: We trained an LCN (softmax regression; ) and an HCN (ResNet-56; ) to classify the colored MNIST dataset.

Techniques:

Journal: Pattern Recognition Letters

Article Title: A too-good-to-be-true prior to reduce shortcut reliance ☆

doi: 10.1016/j.patrec.2022.12.010

Figure Lengend Snippet: Expected effects of the LCN-IWs training on classifying different test cases by the HCN. Correct HCN decisions are in green, incorrect are in red. Training images containing local shortcuts are outlined in magenta. Whereas ordinary (w/o LCN-IWs) training should lead to poor o.o.d. performance on Incongruent test items where the shortcut is now misleading, LCN-IWs should selectively downweight training items with the shortcut allowing the HCN to generalize well across the spectrum. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Article Snippet: We trained an LCN (softmax regression; ) and an HCN (ResNet-56; ) to classify the colored MNIST dataset.

Techniques:

Journal: Pattern Recognition Letters

Article Title: A too-good-to-be-true prior to reduce shortcut reliance ☆

doi: 10.1016/j.patrec.2022.12.010

Figure Lengend Snippet: Accuracies on incongruent, neutral, and congruent test sets after ordinary and HCN-/LCN-weighted training, with (A) local or (B) global shortcuts in training set. HCN is ResNet-56. Across shortcut types, LCN-IWs result in almost equally high accuracy on all sets. HCN-IWs constantly result in accuracies inferior to LCN-IWs; moreover, on neutral and congruent test sets, accuracies after HCN-IWs training are lower than after ordinary training. Thus, LCN-IWs are successful in avoiding shortcut reliance and preserving useful features, while HCN-IWs are not.

Article Snippet: We trained an LCN (softmax regression; ) and an HCN (ResNet-56; ) to classify the colored MNIST dataset.

Techniques: Preserving

Journal: Pattern Recognition Letters

Article Title: A too-good-to-be-true prior to reduce shortcut reliance ☆

doi: 10.1016/j.patrec.2022.12.010

Figure Lengend Snippet: Effects of the LCN-/HCN-IWs training procedure for each of the 45 class pairs depending on a difficulty of the respective binary classification problem. HCN is ResNet-56; (A) local and (B) global shortcut types are considered separately. The effects of training are represented by the Overall Benefit measure ( O B ; gain + loss; see Section 3.2.4 ); the difficulty of a pair is represented by the neutral test accuracy after ordinary training. Recapitulating previous results, LCN-IWs are more effective than HCN-IWs. Furthermore, the easier learning problem, the less O B from IWs: the relatively higher capacity of a network supplying IWs leads to downweighting non-shortcut items.

Article Snippet: We trained an LCN (softmax regression; ) and an HCN (ResNet-56; ) to classify the colored MNIST dataset.

Techniques: