Reddit Users Voted for the Best Papers of 2018 (with Reading Tips)

by Super Neuro

There is no doubt that reading papers is important for gaining a deeper understanding of the technologies and problems in a specific field.

There were also many high-quality papers in 2018, such as award-winning papers at major academic conferences. Today we will review the papers that Reddit users think helped them in 2018.

Reddit user: I vote for this paper

@beezlebub33 Recommended papers:
"Large-Scale Study of Curiosity-Driven Learning"
Research on curiosity-driven learning at scale
https://pathak22.github.io/large-scale-curiosity/

Recommended reasons:
The importance of this article lies in the fact that it achieves excellent performance in many games without a good reward mechanism. What is important is that it learns to play games through prediction, can identify behaviors that violate expectations, and can explore unknown areas. This may be the direction of future development of AI: self-supervision, unlabeled data, prediction, curiosity, internal motivation, etc.

People haven't had enough time to create supervised training sets and define matrices for those data sets, but if you give the AI raw data and it can learn an internal representation of the spatiotemporal evolution of the system, then you can define the goal and eventually use the AI to achieve it.

@YBuzzinGA recommended papers:
"Learning Unsupervised Learning Rules"
Learning unsupervised learning rules
https://arxiv.org/abs/1804.00222.

Recommended reasons:
This post is about using unsupervised learning to accomplish some tasks, which is characterized by the fact that the model is learning how to learn on its own.

Meta-learning is a key area, learning the learning rules that will allow an AI to understand itself and improve itself. If you can teach a computer how to learn to understand itself, then we may be able to make a leap forward.

@breadwithlice recommended papers:
"Phrase-Based & Neural Unsupervised
Machine Translation》
Phrase-based and unsupervised neural machine translation
https://arxiv.org/abs/1810.04805v1

Recommended reasons:
In this article, the translation is completed using only a single corpus, without the need for any mapping, dictionary or parallel data.
The paper uses a reverse translation technique, when converting from A to B, then converting B to A, which greatly improves the translator, and then switching A and B, the result is surprising!

@kartayyar Recommended Papers:
《Pre-training of Deep Bidirectional Transformers
for Language Understanding》
Pre-training of deep bidirectional methods for language understanding
https://arxiv.org/abs/1809.10756

Recommended reasons:
What I like about it:
Great innovative idea, the masking method they used is very creative.

They used very simple sentences to clearly express their core ideas.
There is code on Github to reproduce the results.
Capable of handling a variety of tasks.

@ndha1995 Recommended papers:
"An Introduction to Probabilistic Programming"
Introduction to Probabilistic Programming
https://arxiv.org/abs/1809.10756

Recommended reasons:
This was my favorite paper of 2018.

The author provides a comprehensive and rigorous introduction to probabilistic programming and, in the final chapter, introduces recent research on the combination of deep neural networks and probabilistic programming.

Hardcore advice for reading papers

Even if you know about the great papers, apart from saying wow, they are so great, how can you understand them?

First of all, you should be clear about your motivation. The effect and experience will be completely different if you actively want to explore and passively complete tasks. We found some hardcore suggestions, well, that's all for you.

Read critically

This is a very important attitude. Don't blindly follow the author's point of view. Instead, doubt and verify.

What is critical reading? Try to ask questions. If the author tried to solve a problem, did they solve it correctly? Are there simple solutions that the author didn't consider? What are the limitations of the solution (including those that the author didn't notice or explicitly acknowledge)?

Are the authors' assumptions reasonable? Given the assumptions, is the paper's logic clear and sound, or are there flaws in the reasoning?

If the authors provide data, does their data support their argument? Is the path they took to collect the data reasonable? What about the way they interpret the data? Would other data be better?

Read creatively

Reading a paper critically is not the most difficult, because it is easier to destroy than to build. Creative reading involves harder, more active thinking.

For example: What are the good ideas in this article? Are there other applications or extensions of these ideas? Can they be further generalized? Are there improvements that would make a significant difference? If you were to conduct related research yourself, what would you do next?

Take notes while reading a paper

Many people take notes while reading papers. This is a good idea. Write down any questions or comments that come to your mind in any way you like. Try to find the author's key points.

Mark the most important or seemingly problematic data. This helps in understanding the paper and also helps in reviewing it later.

After the first reading, try to summarize the paper in one or two sentences.

Almost all good papers propose an answer to a specific question. If you can succinctly describe a paper, you probably already understand what the author did, including the question they wanted to answer and the ultimate 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 your paper in one or two sentences, try a different approach and create a three or four bullet point outline that summarizes the main ideas.

If possible, compare the paper to other works

Summarizing a paper is one way to try to determine its scientific contribution. But to really get a handle on the scientific value, you have to compare the paper to other work in the field. Are the ideas novel, or have they been around before?
 
It’s worth mentioning that there are many ways to present scientific research. For example, some papers only propose new ideas, while others implement and verify them and show how they work; others combine previous ideas and integrate them under a novel framework. Knowing other work in the field can help you better understand the value of the paper.