Click the card below to follow the " CVer " public account
AI/CV heavy dry goods, delivered as soon as possible
Fengse is from the concave temple and
reproduced from: qubit (QbitAI)
What can undergraduate students do in scientific research ?
The latest paper included in CVPR 2022 provides a new idea to solve the classification problem of long-tailed distribution data in deep learning , and finally realizes the new SOTA.
Paper: https://arxiv.org/abs/2111.09030
The article has a total of 5 authors, including doctoral students and senior researchers from A*STAR in Singapore...
And in the first place is an undergraduate student from Tianjin University , Li .
The amazing thing is that this is not the first top conference paper of this "newborn calf". Before that, he also won a top conference in the field of data mining (WWW 2022), which is also a work.
Undergraduates engaged in scientific research, has the momentum already been so strong?
If you enter the lab in your junior year, you will have 2 top conferences in your senior year.
Li is from the Department of Intelligence and Computing , Tianjin University, and is a senior this year .
This CVPR article mainly uses a new ensembling learning strategy to solve the problem of long-tail classification.
We all know that the long-tailed distribution is a very common application of deep learning for the classification of these data.
Its difficulty is mainly that the sample size is extremely unbalanced , especially if the tail sample size is too small , it is difficult to obtain effective training results.
Currently, ensemble learning-based methods show great potential, achieving SOTA performance.
But this approach has two limitations:
One is that predictions in failure-sensitive applications are usually unreliable, which has a great impact on tail data, which is very error-prone;
The second is that it allocates a uniform number of resources (experts) to all samples, which results in redundant and high computational costs for simple samples.
Therefore, Li et al. proposed to realize automatic perception of tail category samples by introducing uncertainty ensemble.
On this basis, it is proposed to dynamically allocate more model resources (experts) for tail category samples than head samples to take into account both performance and efficiency.
Finally, the obtained model realizes automatic detection and training adjustment of tail category samples, and becomes a general model for solving long-tail classification problems.
Comprehensive experiments on a series of tasks such as classification, tail detection, outlier detection, and fault prediction demonstrate that the performance of the model successfully beats existing SOTA methods .
In June of this year, Li is about to graduate with an undergraduate degree and will directly enter Purdue University as a doctoral student under the tutelage of Professor Zhang.
Professor Zhang has just graduated from Cornell University with a Ph.D. in Statistics and is an Assistant Professor at Purdue University. Her research interests are machine learning and the construction of probabilistic models in data science.
When he was a junior, he entered the machine learning and data mining laboratory of the faculty with his excellent academic performance .
Under the leadership of Zhang Changqing (Ph.D. supervisor of Tianjin University), Li won the bid for two top papers after just one and a half years of study .
It can be said that the afterlife is terrifying.
More and more undergraduates are starting to send top clubs
However, when it comes to undergraduates engaging in scientific research, in fact, everyone should pay attention. Many schools are becoming more and more open and attaching importance to this matter:
In the junior year or lower , some places will be opened for potential students to enter the laboratory and engage in scientific research with graduate students and doctoral students.
For example, Tsinghua's "Spark Class" recruits sophomore students, Peking University also encourages students to start scientific research projects from their sophomore year, and schools such as Chinese University of Science and Technology will also give special scholarships to undergraduates who have made corresponding achievements .
Under such circumstances, there are already many students like Li who have published the top conference during their undergraduate years, which can be said to be full of halo.
For example , Gao Tianyu, a Tsinghua Special Award winner and a 2016 undergraduate who we are familiar with , has four top conferences in his four years of college: two AAAI, two EMNLP, and a live broadcast dedicated to imparting his own scientific research experience. .
For example , Mao Xiao , an MIT undergraduate student, won the bid for FOCS 2021, the top computer theory conference in his senior year, and won the Best Student Paper Award.
Another example is Liu Hong, an undergraduate student of the Department of Electronics of Tsinghua University, who has 3 papers and a top conference; Wu Kewen, an undergraduate student of Peking University, not only presented the top conference (ACM Annual Conference of Computing Theory STOC), but also won the best paper award; and Wang Tan, an undergraduate student of the University of Electronic Science and Technology of China, also has 1 essay, 1 essay, CVPR 2020, Zhejiang University undergraduate students, 1 essay, 1 essay, ICML 2019...
There are many similar examples.
It can be said that undergraduates are getting more and more powerful in scientific research, what do you think?
Reference link:
http://cic.tju.edu.cn/info/1040/3704.htm
ICCV和CVPR 2021论文和代码下载
后台回复：CVPR2021，即可下载CVPR 2021论文和代码开源的论文合集
后台回复：ICCV2021，即可下载ICCV 2021论文和代码开源的论文合集
后台回复：Transformer综述，即可下载最新的3篇Transformer综述PDF
CVer-Transformer交流群成立
扫描下方二维码，或者添加微信：CVer6666，即可添加CVer小助手微信，便可申请加入CVer-Transformer 微信交流群。另外其他垂直方向已涵盖：目标检测、图像分割、目标跟踪、人脸检测&识别、OCR、姿态估计、超分辨率、SLAM、医疗影像、Re-ID、GAN、NAS、深度估计、自动驾驶、强化学习、车道线检测、模型剪枝&压缩、去噪、去雾、去雨、风格迁移、遥感图像、行为识别、视频理解、图像融合、图像检索、论文投稿&交流、PyTorch、TensorFlow和Transformer等。
一定要备注：研究方向+地点+学校/公司+昵称（如Transformer+上海+上交+卡卡），根据格式备注，可更快被通过且邀请进群
▲扫码或加微信: CVer6666，进交流群
CVer学术交流群（知识星球）来了！想要了解最新最快最好的CV/DL/ML论文速递、优质开源项目、学习教程和实战训练等资料，欢迎扫描下方二维码，加入CVer学术交流群，已汇集数千人！
▲扫码进群
▲点击上方卡片，关注CVer公众号
整理不易，请点赞和在看