# Paper Review

{%hackmd D5Ke1wgzREKAMTq40fd_Ww %}

## 2022

### Improving clustering uncertainty-weighted embeddings for active domain adaptation.

* Use proposed new Active DA framework ADA-DWUS to address the issue that uncertainty as estimated by the classifier suffers from large variance, leading to unreliable queries and thus uninformative instances.
* My opinion: This paper combine active learning and domain adapation. Despite the fact that I am not proficient in AL and DA, but maybe I can try to understand more about them.
* interest index(1~5): 3

### Reduction from complementary-label learning to probability estimates

* Propose a new method to improve one of weakly supervised learning - complementary-label learning(CLL).
* My opinion: One possible way to accelerate the process of training, I would like to study more about that.
* interest index(1~5): 4

## 2021

### Improving Model Compatibility of Generative Adversarial Networks by Boundary Calibration

* Use Boundary-Calibration GANs (BCGANs) to solve model compatibility problem or machine learning efficacy.
* My opinion: Although not related to AutoML but this topic sounds interesting.
* interest index(1~5): 3.5

### On the Role of Pre-training for Meta Few-Shot Learning

* Study about the pre-training from the disentanglement of the representations.
* My opinion: This work tries to combine concepts of transfer learning and meta learning on few-shot learning in order to speed-up episodic training(reduce the number of episodes). A bit like what I want to do on training-free learning.
* interest index(1~5): 4

### Active Refinement for Multi-Label Learning: A Pseudo-Label Approach

* A approach to use pseudo label that not only assigns labels to unknown entries but also enjoys significantly better performance than other methods for learning fine-grained classifiers under the limited supervision setting.
* My opinion: A research topic about multi-label learning and weakly supervised. I need to learn more to understand the meaning behind the words but it seems like a fantasy problem to extend.
* interest index(1~5): 3.5

### ADAPTIVE AND GENERATIVE ZERO-SHOT LEARNING

* Yield virtual classes and data by mixup interpolations. Beside, introduce the concept of representing each class with image-adaptive semantic features that could vary among intra-class samples.
* My opinion: This work is related to GZSL, trying to address the issue about predicting the class label of unseen catagory. Though it doesn't resemble what I mentioned yesterday, it seems like that I am disgusted with those revelant topics.
* interest index(1~5): 3

## 2020

### Cost Learning Network for Imbalanced Classification

* Combine cost-sensitive and reinforcement learning to tackle the problem of imbalanced classification.
* My opinion: not mentioned in AutoML but it seems like interesting.
* interest index(1~5): 3.5

### Improving Unsupervised Domain Adaptation with Representative Selection Techniques

* Bio-chemistry application requires considering both the covariate shift and label shift properly.
* My opinion: UDA application seems like a big topic in deep laerning. I am not sure that it can be used to speed up the training process, maybe I need to know more about that.
* interest index(1~5): 3

### My interest rank

| interest index(1~5) | topic |
|:-------------------:|:----------------------------------------------------------------------------------------------------:|
| 5 | AutoML</br> training-free deep learning</br> neural archtecture search</br> |
| 4 | active learning</br> weakly-supervised</br> few-shot learning</br> complementary-label learning</br> |
| 3.5 | Model Compatibility</br>Multi-Label Learning</br> |
| 3 | cost-sensitive learning</br> domain adaptation</br> zero-shot learning</br> |
| 2 | NULL |
| 1 | NULL |