当前位置:网站首页>【学术相关】为什么很多国内学者的AI的论文复现不了?
【学术相关】为什么很多国内学者的AI的论文复现不了?
2022-07-29 07:59:00 【51CTO】
吴恩达老师曾经说过,看一篇论文的关键,是复现作者的算法。
然而,很多论文根本就复现不了,这是为什么呢?
一、数据关系
因为作者使用的数据比较私密,一般人拿不到,这种情况下,即使作者提供了源代码,但是读者却拿不到数据,也就没法复现算法。
这种情况在国内学术界很普遍,数据别人没有,这就好像一位奥数老师,自己出了一道奥数题,自己解答出来,然后把解题过程写了论文,这类论文往往说服力不够,故事性不够强。
二、硬件原因
深度学习的很多算法,是靠大力出奇迹的方法做出来的。比如谷歌、facebook的一些算法,依靠强大的硬件训练出来。
普通研究者没有那么强大的硬件资源,估计达不到他们的1%的算力,根本无法复现算法。
三、数据划分和训练方式
有些论文公开了代码,也公开了数据,但是论文里没有提到数据划分问题,数据如果比较少的话,不同的划分会导致结果不同。
四、众所周知的原因
这个原因大家心知肚明,我就不说太明白了,这个情况出现在很多国内作者的论文里。这个在公开数据上比较少见。
很多国内的学者发的论文,通常的套路是:
1. 定义一个很新但是意义不大的问题;
2. 面向github编程;
3. 网络中加一些attention,module,normalization,loss,加到不会崩为止;
4. 编故事,写小说,看上去逻辑挺清晰的,但是不给别人复现的机会。
理想的论文什么样?
1.效果可以复现,论文里每一个实验逻辑都很通畅,论文所有实验形成的逻辑链完备,使用公开数据集,效果和论文基本一致。
能够达到这种的一般是领域内的大佬,比如陈天奇、何恺明。
2.使用公开数据,公开代码,论文细节清楚,能复现论文的效果。尽管很多论文的作者也不能说明为什么这么设计的网络效果好,这个应该是深度学习可解释性差的原因。因为他们公开了代码,在公开数据集上效果好,能复现效果,所以也是好论文。
文末推荐
最后,推荐一个网站:paperswithcode.com/,很多优秀论文的代码都能找到。
边栏推荐
- Mutationobserver document learning
- After the access database introduces DataGridView data, an error is displayed
- Write some DP
- Chaos and future of domestic digital collections
- Cs61abc sharing session (VI) detailed explanation of program input and output - standard input and output, file, device, EOF, command line parameters
- UPC little C's King Canyon
- [cryoelectron microscope | paper reading] interpretation of sub fault average m software: multi particle cryo EM refining with M
- Shell script - global variables, local variables, environment variables
- [lecture notes] how to do in-depth learning in poor data?
- Monitor the bottom button of page scrolling position positioning (including the solution that page initialization positioning does not take effect on mouse sliding)
猜你喜欢
Analyze the roadmap of 25 major DFI protocols and predict the seven major trends in the future of DFI
Monitor the bottom button of page scrolling position positioning (including the solution that page initialization positioning does not take effect on mouse sliding)
MySQL 45 | 08 is the transaction isolated or not?
[cryoelectron microscope] relation4.0 - subtomogram tutorial
NLP introduction + practice: Chapter 5: using the API in pytorch to realize linear regression
Technology sharing | quick intercom integrated dispatching system
10 practical uses of NFT
2022 Shenzhen Cup Title A: get rid of "scream effect" and "echo room effect" and get out of the "information cocoon room"
How to connect VMware virtual machine to external network under physical machine win10 system
10 common software architecture modes
随机推荐
Ionicons icon Encyclopedia
In JS, 0 means false, and non-0 means true
UE4 principle and difference between skylight and reflecting sphere
Basic introduction to pod
330. Complete the array as required
QT connects two qslite databases and reports an error qsqlquery:: exec: database not open
CentOS deploy PostgreSQL 13
Output 1234 three digits without repetition
[skill accumulation] presentation practical skill accumulation, common sentence patterns
Effective learning of medical image segmentation annotation based on noise pseudo tags and adversarial learning
Cfdiv1+2-bash and a high math puzzle- (gcd+ summary of segment tree single point interval maintenance)
In an SQL file, a test table and data are defined above, and you can select* from the test table below
An Optimal Buffer Management Scheme with Dynamic Thresholds论文总结
Mutationobserver document learning
Dilworth theorem
postman接口测试|js脚本之阻塞休眠和非阻塞休眠
[paper reading | cryoelectron microscope] interpretation of the new subtomogram averaging method in relion 4.0
Joseph Ring problem
The database uses PSQL and JDBC to connect remotely and disconnect automatically from time to time
MapReduce steps of each stage