问题阐述作者在实验中发现,如果让网络在基类训练时保持“变换不变性(invariant)”,比如旋转不变性,平移不变性等,网络对特征的提取、概括能力会增强,但是泛化到新类上的性能会下降;如果让网络在基类训练时保持“等变性(equivariant)”,也就是说网络知道变换后的图片和变换之前的图片之间发生的是什么变换(对不同的变换形式,比如旋转、平移等可以分类辨别),泛化到新类上的能力就会增强。1. However, recently it has been shown in [56] that the quick adaptation of the base learner crucially depends on feature reuse. 然而,最近在[56]中已经表明,基础学习器的快速适应至关重要地依赖于特性重用。[56] Aniruddh Raghu, Maithra Raghu, Samy Bengio, and OriolVinyals. Rapid learning or feature reuse? towards understanding the effectiveness of maml. In International Conference on Learning Representations, 20202. This brings in an interesting question: How much further can FSL performance be pushed by simply improving the base feature extractor? 这带来了一个有趣的问题:通过简单地改进基本特征提取器,FSL的性能能提高多少?3. The generalization of these algorithms often relies on the effective design of inductive biases, since they encode our priori preference for a particular set of solutions. 这些算法的推广通常依赖于归纳偏差的有效设计,因为它们编码了我们对特定解集的先验偏好。4. However, such inductive biases do not provide optimal generalization on FSL tasks and the design of efficient inductive designs for FSL is relatively unexplored. 然而,这种归纳偏差并不能提供FSL任务的最优泛化,并且FSL的高效归纳设计的设计也相对未被探索。
03
方法和细节
基础
1. 不使用episodic training(meta-training):However, following recent works [73, 15, 10], we do not use an episodic training scheme which allows us to train a single generalizable model that can be efficiently used for any-way, any-shot setting without retraining. 2. 使用基类数据简单训练基础特征提取器,CNN+分类器+CE Loss:we train our base learner on the whole base training set Db in a supervised manner. The role of this base learner is to extract good feature embeddings that can generalize for novel classes. 3. 使用了L2正则项
主要创新点:Injecting Inductive Biases through SSL
We propose to enforce equivariance and invariance to a general set of geometric transformations T by simply performing self-supervised learning (SSL).虽然作者的故事讲的很复杂,但是方法很简单:既然网络的两种性质“invariant”和“equivariant”有各自的长处和短处,那就想办法把他们结合起来。用的是图像类别损失(最基础的损失,本质是CE Loss)+图像变换类别损失(强化“equivariant”, 本质是CE Loss)+ 对比损失(强化”invariant“, 本质是infoNCE Loss),见下图