When Deep Classifiers Agree: Analyzing Correlations between Learning Order and Image Statistics

TL;DR

This paper investigates how deep classifiers tend to agree on the order of data instances learned, linking this phenomenon to dataset statistics across various datasets and architectures.

Contribution

It introduces a metric to quantify classification agreement over time and demonstrates its correlation with dataset image statistics across multiple datasets.

Findings

Agreement is consistent across different architectures and hyper-parameters.

Classification order correlates with core image statistics.

Agreement phenomenon is dataset-dependent, not architecture-dependent.

Abstract

Although a plethora of architectural variants for deep classification has been introduced over time, recent works have found empirical evidence towards similarities in their training process. It has been hypothesized that neural networks converge not only to similar representations, but also exhibit a notion of empirical agreement on which data instances are learned first. Following in the latter works$'$ footsteps, we define a metric to quantify the relationship between such classification agreement over time, and posit that the agreement phenomenon can be mapped to core statistics of the investigated dataset. We empirically corroborate this hypothesis across the CIFAR10, Pascal, ImageNet and KTH-TIPS2 datasets. Our findings indicate that agreement seems to be independent of specific architectures, training hyper-parameters or labels, albeit follows an ordering according to image statistics.