Summary
This paper proposes solve the meta-learning problem by taking augmented data samples as the query set during training phase. Generally, the writing of paper is quite poor, the proposed method is filled with subjective conjectures and confusing motivations. I do not think that the proposed method makes sense.
Strengths
- The experimental results looks good.
Weaknesses
__Main concerns:__
- __[Poor writing]__ The writing of this paper is poor. For example, the format of references are not well managed. More efforts are required to improve the quality of the paper.
- __[Unclear formulation]__ The problem studied in this paper is not well formulated, and the notation used in this paper is quite confusing.
- __[Farfetched statements]__ There are many confusing statements that are confusing and farfetched in this paper.
- __[Inconvincing results]__ Due to the confusing method settings, I do not think the empirical results are convincing.
__Minor concerns:__
- __[Limited Novelty]__ The proposed method is not novel.
Questions
1. (Line 44.) Why large datasets where the probability of drawing samples from the same class is much lower helps achieve better performance?
2. (Line 119.) I do not think that $x_{i, j}$ is an appropriate way to represent data samples. Besides, we usually use the upper case to denote the variables, it is a little bit confusing to use it to represent data pool. Meanwhile, if $X_N$ denotes the unlabelled samples, how to sample $n$-way $1$-shot tasks?
3. (Line 121.) What does the sentence "We only design n-way (n is the number of ways or classes), 1-shot support
122 sets because each sample in the support set is drawn randomly, and we cannot randomly add more
123 same-class support samples to that set" mean? In meta-learning settings, $N$-way $K$-shot tasks, such as 5-way 5-shot tasks, are common.
4. (Line 127.) What does $y_{i, j}$ mean? What the function of so-called "labeled values"? Do you mean the random assigned class numbers?
5. What does the notation $f(A)$ mean? What is $A$?
6. In my opinion, the query set in meta-learning mean the data belonging to the same classes as support data but are not observed by model during adaptation phase (inner loop). So I suggest that use other term to describe the augmented support data to make it clear.
7. In Line 128, $c$ means the number of ways, but it is defined further in Line 136 as the total number of classes. So, what is the structure of the tasks?
8. Line 134. What does the sentence "We intuitively know that when we draw a few samples from a large pool of data, more than one sample belonging to the same class is low" mean?
9. In Line 139-140, the authors claim that Omniglot has less classes than mini-ImageNet. In fact, mini-ImageNet only owns 100 classes with 600 samples in each class. Besides, why is it more likely that all drawn samples will originate from different classes for datasets that have less number of samples in each class? In fact, I do not understand the generation process of support-query set described in Line 134-146.
10. Sometimes augmentations tend to modify the information in images, so how do these techniques help improve the performance?
11. What are the evidence for your statement that the classifier of MAML trained on RGB samples has a severe overfitting issue ...... Therefore, the classifier learns very little during the training phase? Such statement is farfetched.
12. The expressions of $\theta_i^{'}$ and $\theta^i$ are inconsistent.
13. Where is the semi-supervised meta-learning algorithm? What the goal of section 3.2?
__Small problems:__
1. Line 139. It would be better to replace "prior" with "former".
Rating
3: reject, not good enough
Confidence
4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.