No Reason for No Supervision:
Improved Generalization in Supervised Models
ICLR 2023 (spotlight, notable top 25%)
Mert Bulent Sariyildiz1,2, Yannis Kalantidis1, Karteek Alahari2, Diane Larlus1
1 NAVER LABS Europe 2 Inria
ImageNet-1K (IN1K) vs transfer task performance for ResNet50. We report IN1K (Top-1 accuracy) and transfer performance (log odds) averaged over 13 datasets (the 5 ImageNet-CoG concept generalization datasets, Aircraft, Cars196, DTD, EuroSAT, Flowers, Pets, Food101 and SUN397) for a large number of our models trained with the supervised training setup we propose. Models on the convex hull are denoted by stars. We compare to public state-of-the-art (SotA) models: the supervised RSB-A1 and SupCon models, the self- supervised DINO, the semi-supervised PAWS and a variant of LOOK using multi-crop.
We consider the problem of training a deep neural network on a given classification task, e.g., ImageNet-1K (IN1K), so that it excels at both the training task as well as at other (future) transfer tasks. These two seemingly contradictory properties impose a trade-off between improving the model’s generalization while maintaining its performance on the original task. Models trained with self-supervised learning tend to generalize better than their supervised counterparts for transfer learning; yet, they still lag behind supervised models on IN1K. In this paper, we propose a supervised learning setup that leverages the best of both worlds. We extensively analyse supervised training using multi-scale crops for data augmentation and an expendable projector head, and reveal that the design of the projector allows us to control the trade-off between performance on the training task and transferability. We further replace the last layer of class weights with class prototypes computed on the fly using a memory bank and derive two models: t-ReX that achieves a new state of the art for transfer learning and outperforms top methods such as DINO and PAWS on IN1K, and t-ReX* that matches the highly optimized RSB-A1 model on IN1K while performing better on transfer tasks. Finally, we perform several analyses of the features and class weights to present insights on how each component of our setup affects the training and learned representations.
Citation
@inproceedings{sariyildiz2023improving,
title={No Reason for No Supervision: Improved Generalization in Supervised Models},
author={Sariyildiz, Mert Bulent and Kalantidis, Yannis and Alahari, Karteek and Larlus, Diane},
booktitle={International Conference on Learning Representations},
year={2023}
}