Trivial or impossible -- dichotomous data difficulty masks model differences (on ImageNet and beyond)

TL;DR

This paper investigates how dichotomous data difficulty in datasets like ImageNet influences model decisions, revealing that most images are either trivial or impossible, which masks differences between models and impacts understanding of their biases.

Contribution

It demonstrates that dichotomous data difficulty dominates model decision boundaries, and removing trivial and impossible images reveals clearer model differences.

Findings

46.0% of images are trivial for models

11.5% of images are impossible beyond label errors

Humans predict image difficulty for CNNs with 81.4% accuracy

Abstract

"The power of a generalization system follows directly from its biases" (Mitchell 1980). Today, CNNs are incredibly powerful generalisation systems -- but to what degree have we understood how their inductive bias influences model decisions? We here attempt to disentangle the various aspects that determine how a model decides. In particular, we ask: what makes one model decide differently from another? In a meticulously controlled setting, we find that (1.) irrespective of the network architecture or objective (e.g. self-supervised, semi-supervised, vision transformers, recurrent models) all models end up with a similar decision boundary. (2.) To understand these findings, we analysed model decisions on the ImageNet validation set from epoch to epoch and image by image. We find that the ImageNet validation set, among others, suffers from dichotomous data difficulty (DDD): For the range of investigated models and their accuracies, it is dominated by 46.0% "trivial" and 11.5% "impossible" images (beyond label errors). Only 42.5% of the images could possibly be responsible for the differences between two models' decision boundaries. (3.) Only removing the "impossible" and "trivial" images allows us to see pronounced differences between models. (4.) Humans are highly accurate at predicting which images are "trivial" and "impossible" for CNNs (81.4%). This implies that in future comparisons of brains, machines and behaviour, much may be gained from investigating the decisive role of images and the distribution of their difficulties.