No One Representation to Rule Them All: Overlapping Features of Training Methods

Abstract

Despite being able to capture a range of features of the data, high accuracy models trained with supervision tend to make similar predictions. This seemingly implies that high-performing models share similar biases regardless of training methodology, which would limit ensembling benefits and render low-accuracy models as having little practical use. Against this backdrop, recent work has developed quite different training techniques, such as large-scale contrastive learning, yielding competitively high accuracy on generalization and robustness benchmarks. This motivates us to revisit the assumption that models necessarily learn similar functions. We conduct a large-scale empirical study of models across hyper-parameters, architectures, frameworks, and datasets. We find that model pairs that diverge more in training methodology display categorically different generalization behavior, producing increasingly uncorrelated errors. We show these models specialize in subdomains of the data, leading to higher ensemble performance: with just 2 models (each with ImageNet accuracy ~76.5%), we can create ensembles with 83.4% (+7% boost). Surprisingly, we find that even significantly low-accuracy models can be used to improve high-accuracy models. Finally, we show diverging training methodology yield representations that capture overlapping (but not supersetting) feature sets which, when combined, lead to increased downstream performance.

References in corpus (21)

• Very Deep Convolutional Networks for Large-Scale Image Recognition
• An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale
• Learning Transferable Visual Models From Natural Language Supervision
• Popular Ensemble Methods: An Empirical Study
• What Uncertainties Do We Need in Bayesian Deep Learning for Computer Vision?
• On Calibration of Modern Neural Networks
• Scaling Up Visual and Vision-Language Representation Learning With Noisy Text Supervision
• AutoAugment: Learning Augmentation Policies from Data
• DenseNet: Implementing Efficient ConvNet Descriptor Pyramids
• Do ImageNet Classifiers Generalize to ImageNet?
• Adversarial Examples Are Not Bugs, They Are Features
• Deep Ensembles: A Loss Landscape Perspective
• What is being transferred in transfer learning?
• Hyperparameter Ensembles for Robustness and Uncertainty Quantification
• Dual Path Networks
• Beyond accuracy: quantifying trial-by-trial behaviour of CNNs and humans by measuring error consistency
• Mitigating Bias in Calibration Error Estimation
• The Evolution of Out-of-Distribution Robustness Throughout Fine-Tuning
• Fix your classifier: the marginal value of training the last weight layer
• Model Similarity Mitigates Test Set Overuse
• Do Self-Supervised and Supervised Methods Learn Similar Visual Representations?