https://mp.weixin.qq.com/s/dQxbmSHouyAx13LxDKl8kQ
by 超神经
读论文这件事，对深入了解细分领域的技术和难题来说，重要性毋庸置疑。
2018 年也出现了很多优质的论文，比如各大顶级学术会议上的获奖论文，我们今天盘点的是 Reddit 网友心中那些 2018 帮助过他们的论文。Reddit netizen: I vote for this paper
@beezlebub33 Recommended paper:
"Large-Scale Study of Curiosity-Driven Learning"
Large-scale curiosity-driven learning research
https://pathak22.github.io/large-scale-curiosity/
Reason for recommendation:
The importance of this article is reflected in its excellent performance in many games without a good reward mechanism. The important thing is that it learns to play games through predictions, can identify behaviors that violate expectations, and can explore unknown areas. This may be the future direction of AI: self-monitoring, unlabeled data, predictability, curiosity, internal motivation, etc.
People have not had enough time to create supervised training sets and define matrices for these data sets. But if you provide AI with raw data, and it can learn the internal representation of the system's spatiotemporal evolution, then you can define goals and finally achieve it with AI.
@YBuzzinGA Recommended paper:
"Learning Unsupervised Learning Rules"
learning unsupervised learning rule
https://arxiv.org/abs/1804.00222 .
Recommended reason:
This article is about using unsupervised learning to complete some tasks. Its characteristic is that the model is learning how to learn by itself.
Meta-learning is a key area. Learning those learning rules will allow AI to understand itself and improve itself. If you can teach a computer how to learn to understand itself, then we may make a leap.
@breadwithlice recommended paper:
"Phrase-Based & Neural Unsupervised
Machine Translation"
based on phrase and unsupervised neural machine translation
https://arxiv.org/abs/1810.04805v1
Reason for recommendation: In
this article, only a single corpus can be used to complete the translation without any mapping, dictionary or parallel data.
A reverse translation technique is used in the paper. When converting from A to B, then converting B to A. This greatly improves the translator, and then switching A and B. This result is amazing!
@kartayyar recommended paper:
"Pre-training of Deep Bidirectional Transformers
for Language Understanding " Pre-training of the
deep bidirectional method in language understanding
https://arxiv.org/abs/1809.10756
Recommended reason:
What I like about it:
great innovative concept, the masking method they use is very creative.
They clarified their core philosophy in very simple sentences.
There are code reproducible results on Github.
Able to handle a number of different tasks.
@ ndha1995 Recommended paper:
"An Introduction to Probabilistic Programming"
probabilistic planning Introduction
https://arxiv.org/abs/1809.10756
Recommended reason:
This is my favorite in the 2018 paper.
The author gave a comprehensive and rigorous introduction to probabilistic planning, and in the last chapter, introduced the recent research on the combination of deep neural networks and probabilistic planning.
Hardcore suggestions for reading papers
Even if you know great papers, except for shouting wow, which are really great, how can you understand them?
First of all, think about your motivation. Actively want to explore and passively complete tasks. The effect and experience will be completely different. We found some hardcore suggestions, um, we will help you here.
Read critically
This is a very important attitude. Don't blindly follow the author's point of view. Instead, go to doubt and verify.
What is critical reading? Try to ask questions. If the author tried to solve a problem, did they solve it correctly? Is there an author who did not consider a simple solution? What are the limitations of the solution (including those that the author did not notice or explicitly acknowledged)?
Is the author's assumption reasonable? Given the hypothesis, is the logic of the paper clear and reasonable, or is there a flaw in the reasoning?
If the authors provide data, can their data be used to confirm their argument, and is their path to collecting data reasonable? How do they interpret the data? Would it be better to switch to other data?
Read creatively
It is not the most difficult to read a paper critically, because it is easier to destroy than to construct. And creative reading involves harder, more positive thinking.
For example: What are some good ideas in this article? Are there other applications or extensions to these ideas? Can they be further promoted? Are there improvements that would make a major difference? If you do the relevant research yourself, what will you do next?
Take notes while reading the paper
Many people take notes while reading papers. This method is very good. Record any questions or comments you think of in the way you like. Try to find the key points of the author.
The highly acclaimed Cornell method
Mark the most important or seemingly problematic data. Such a mark helps to understand the paper, but also helps to review later.
After the first reading,
try to summarize the paper in one or two sentences
Almost all excellent papers are asking the answer to a specific question. If you can describe a paper succinctly, you may have understood the author’s work, including the problem they want to solve and the final answer. Once you focus on the main idea, go back and outline the paper to get a deeper understanding of the specific details.
In fact, if it’s easy to summarize the paper in one or two sentences, you can try another method and list three or four outlines to summarize the main ideas.
If possible,
compare the paper with other works
Summarizing the paper is a way to try to determine the scientific contribution of the paper. But to truly grasp the scientific value, you must compare the paper with other works in the field. To figure out whether these ideas are new or have appeared before?
It is worth mentioning that there are many ways to present scientific research. For example, some papers only propose new ideas, while others implement verification and show how they work; others combine previous ideas and integrate them under a novel framework. Understanding other work in the field can help you better understand the value of the paper.
Super Nervous Encyclopedia
Unsupervised learning
Machine learning can be divided into supervised learning, unsupervised learning and semi-supervised learning.
Unsupervised learning is a learning method that does not provide corresponding category identification for the training set. It is suitable for situations where there is a data set but no labels.
The data used in unsupervised learning is unlabeled, and the output result corresponding to the input data is not known during training. No supervised enhancement (tell it what kind of learning is correct).
Unsupervised learning can only read the data silently, looking for data models and laws, such as clustering (grouping similar data into a group) and anomaly detection (looking for anomalies).
The purpose of unsupervised learning is to classify the original data in order to understand the internal structure of the data. Its characteristic is that it only provides input examples for this kind of network, and it automatically finds its potential category rules from these examples. After learning and testing, you can also apply it to new cases.
