NeurIPS 2019
Sun Dec 8th through Sat the 14th, 2019 at Vancouver Convention Center
Paper ID:2213
Title:Cross Attention Network for Few-shot Classification

Reviewer 1

Post rebuttal ========== I’d like to thank the authors for performing the additional ablation, comparisons via visualizations, and experiment in a cluttered environment, as I suggested in my reviews. I think these additional results would be good additions (to the Appendix at the very least) and strengthen the paper. I continue to recommend acceptance. I do agree with R3 though that the proposed transductive method is very similar to previous works for semi-supervised learning, and it would be useful to be more clear about this in the writing. Before rebuttal ============ Summary This paper introduces a state-of-the-art approach to few-shot classification. There are two orthogonal components proposed: the first influences the embedding function applied to the images of an episode, and the second introduces a strategy for using the query set of each episode in a transductive manner as additional unlabeled data for refining the within-episode classifier. The insight behind their embedding method is that it is useful to process each query example differently in light of the given support set, instead of processing each example of the episode independently. The motivation is that this may allow to identify objects that are common between a query example and the support set which are likely to be the target objects that the classification decision should be based on. Otherwise, it may be the case that the feature extraction emphasizes the identification of objects that are in the meta-training set, even if they are background (not target objects) for a particular episode. Within each episode, they first compute a prototype per class by averaging the (standard) embeddings of the support examples of each class. They also embed each query example independently to start with. Then, in the embedding space of the last convolutional layer, they compute for every prototype p and query example q, the channel-wise cosine similarity at each of the height x width spatial locations of that layer. This yields an m x m correlation matrix for the specific pair of class prototype and query example, where m = height x width. This correlation matrix is used for determining an attention map (over the spatial locations) for the processing of the prototype and one for the processing of the query example. These maps are obtained via a m x 1 convolution operation over the rows and columns of the correlation matrix, respectively. This convolutional layer is referred to as the meta-fusion layer and is meta-learned (i.e. updated based on the loss on the query set of each meta-training episode). Meta-learning this layer encourages the resulting attention mechanism to emphasize those features that are most discriminative for the classification task at hand. Once the attention maps are obtained, the prototype and the query example are re-embedded using the updated feature maps. The loss of the episode is then the nearest-neighbor based classification of the query examples. This episode-specific loss is combined with a global classification loss (over all classes of the training set) in a joint training fashion. Their proposed transductive strategy, in short, consists in treating the most confident (unlabeled) query examples of each episode as additional labeled examples that can be used to compute the class prototypes. The threshold that the confidence level of a query point must surpass in order to contribute to prototype creation becomes more permissive throughout training via a manually-defined schedule. This method achieves state-of-the-art on common datasets. Further, they run some ablations since their method involves a few components that are orthogonal to the proposed architecture and are not standard, namely joint training, and transductive treatment of the query set. For measuring the effect of the former, they experiment with a standard Prototypical Network and with a Prototypical Network that was jointly trained with the usual episodic query loss as well as a global classification via a softmax layer with as many outputs as total training classes (in the same way as they train their method). The fact that the latter outperformed the former significantly indicates that this process may be useful more generally for training meta-learning models. For assessing the benefit of their transductive approach, they applied it on other models too and showed that they also notably benefit from it, which is also an interesting result. Comments 1) I really like the type of exploration done in Figure 1 where the authors attempt to visualize which parts of the image their proposed method (and a baseline) pay attention to. I would argue that simply looking at classification accuracy results is not sufficient to conclude that the reason for improvement witnessed here is solely due to addressing the identified weakness (since unfortunately there are other confounding factors). Further, orthogonally to the accuracy results, it is an interesting finding if standard approaches indeed suffer from this and the proposed method provides a remedy. I would therefore focus on these qualitative results more, and explain in the main text (not just the appendix) exactly how those visualization are created, and show those results for various models. 2) Somewhat related to the previous point: Pure metric-based models like Prototypical Networks lack an explicit mechanism for adaptation to each task at hand and it therefore seems plausible that they indeed suffer from the identified issue. However, it is less clear whether (or to what extent) models that do perform task-specific adaptation run the same danger. Intuitively, it seems that task adaptation also constitutes a mechanism for modifying the embedding function so that it favours the identification of objects that are targets of the associated classification task. By task adaptation here I’m referring either to gradient-based adaptation (as in MAML and variants) or amortized conditioning-based adaptation (as in TADAM for example). Therefore, it would be very interesting to empirically compare the proposed method to these other ones not only in terms of classification accuracy but also qualitatively via visualizations as in Figure 1 that show the areas of the image that a model focuses more for making classification decisions. 3) Suggestion for the transductive framework: In Equation 8, it might be useful to incorporate the unlabeled examples in a weighted fashion instead of trusting that every example whose confidence surpasses a manually-set threshold can safely contribute to the prototype of the class that it is predicted to belong to. Specifically, the contribution of an unlabeled example to the updated class prototype can be weighted by the cosine similarity between that unlabeled example and that prototype (normalized across classes) and maybe additionally by the confidence c_b^q. This might slightly relieve the need to find the perfect threshold, since even if it is not conservative enough, a query example will be prohibited by modifying a prototype too much. An example of this is in Ren et al. [1] when computing refined prototypes by including unlabeled examples. 4) It seems that the weakness that this method is addressing would be more prominent in images comprised of multiple objects, or cluttered scenes. It would be very interesting to compare this approach to previous ones on few-shot classification on such a dataset! 5) For more easily assessing the degree of apples-to-applesness of the various comparisons in the tables, it would be useful to note which of the listed methods use data augmentations (as until recently this was not common practice for few-shot classification), what architecture they use, and what objective (most are episodic only but I think TADAM also performs joint training as the proposed method). 6) Another difference between the proposed approach and previous Prototypical Network-like methods is that the distance comparisons that inform the classification decisions are done in a feature-wise manner in this work. Specifically, when comparing embeddings a and b, for each spatial location, the distance between the feature vectors of a and b at that location is computed. The final estimate of the distance between a and b is obtained by aggregating those feature-wise distance estimates over all spatial locations. In contrast, usually the output of the last embedding layer is reshaped into a single vector (of shape channels x height x width) and distance comparisons of examples are made by directly comparing these vectors. It would therefore be useful to perform another ablation where a standard Prototypical Network is modified to perform the same type of distance comparison as their method. 7) Similarly to how the proposed transductive method was applied to other models, it would be nice to see results where the proposed joint training is also applied to other models, since this is orthogonal to the choice of the meta-learner too. References [1] Meta-Learning for Semi-Supervised Few-shot Classification. Ren et al. ICLR 2018.

Reviewer 2

This paper presents cross-attention for few-shot learning. It performs cross-correlation between the query image and each image in the support set. The obtained cross-attention maps are then used to gate the feature maps to obtain the final feature maps from both query and support images. I think the motivation is sound and proposed cross-attention maps for query and support images are novel. The experimental results validates the effectiveness of the proposed method. My major concern is on the time complexity of the proposed method, as it requires to conduct cross-correlation between query and every support images. In the comparision table, I think the authors should consider time complexity of different compared methods. I think the idea is novel and results show SOTA performance. My only concern is the computational complexity, as it needs to

Reviewer 3

Post Rebuttal: Thank you to the authors for their comments. I still believe that the discussion of the proposed transductive method should cite previous work in self-training methods for semi-supervised learning, as this is essentially what the authors have proposed for taking advantage of unlabeled query data. I am happy that the authors are committed to releasing the code for their method. Before Rebuttal: Summary This paper proposes a modification of embedding-based few-shot learning methods, where instead of embedding support and query set items independently, each prototype embedding is computed conditional on a query item and each query item is computed conditional on a prototype. Specifically, the conditional embedding takes the form of an attention module which is computed using a correlation layer and a fusion layer. The idea is that this type of attention allows the embedding to highlight specific objects with respect to another input. Classification is done via distance computed using this conditionally-computed embeddings and additionally a global classification layer, which classifies each query set item to a class label in the training set independent of the current episode. Lastly, a method for transductive inference is proposed, which involves taking the top-t classified query items and adding them to the support set with the predicted label and re-calculating the embeddings and repeating for a certain number of iterations. The method is evaluated on the mini-Imagenet and tiered-Imagenet benchmarks. Additional ablation studies are conducted to study the effects of different parts of the proposed model. Strengths 1. Paper is well-written and describes the proposed model well. 2. Experiments involve more than just pure results, as ablation studies validate different parts of model. 3. Achieves state-of-the-art results for Mini-Imagenet and Tiered-Imagenet (larger margin on Tiered-Imagenet) Weaknesses 1. Proposed model involves a lot of complicated moving parts - it is not clear whether it'll be easily reproducible given that it is so complicated. 2. I don't believe the proposed transductive method is very novel as I believe its related to a common way to incorporate unlabeled data in semi-supervised methods (see self-training methods in semi-supervised learning).