How to effectively read research papers? — For beginners

It’s key to decide what to ‘not’ read and what to read!

AI/Machine Learning technologies have been evolving very rapidly today, and actually this is not the only trend we can see from those specific fields — in general, all sciences and engineering technologies are changing so fast to the level we can’t easily catch up what’s really happening. This is probably because we’re living in the age of society that invests a large amount of resources in those advances to potentially take a financial, social benefit from the outcomes.

Under such a flood of information and knowledge, you might often be thinking about what would be an effective strategy to keep track of this fast-paced development without being left behind. Even if you’re really passionate about AI and Machine Learning, however, I think numerous works newly introduced from scientific journals, conferences, social media, and blogs every day could overwhelm you very easily.

Therefore, I’d like to share with you through several posts how I’m dealing with this challenge as a person who is working in academia, and in this post, and probably in a couple of more later, I’ll show you my approaches to learning scientific findings from published papers — Effective reading of research papers. Although I’ll take an AI-related research paper as an example below because I have a background in it, I believe this approach could generally be applied to the reading in any other STEM fields. A disclaimer is, however, what I propose here is simply based on my personal research experiences, and if you don’t feel it like a good solution for you, I’d recommend you google “How to read academic papers” to find the ones to better fit your style.

Decide what to “not” read according to your goal and familiarity

For effective paper reading, it is very important to first define “your goal” and “your familiarity” where “your familiarity” here refers to “how much you are familiar with the field in which the paper is situated”.

In particular, this post discusses reading activity as a “beginner”, and here are the definitions accordingly:

• Beginner: Reader who isn’t familiar with the field of the paper
• Goal: Not to ‘understand 100% of the paper immediately’ but to ‘get a rough idea about the field and what the authors have done’

Would this definition matter in reading? Yes, because paradoxically it helps you select the parts not to read. Each research paper generally consists of multiple sections such as Title, Abstract, Introduction, Related Work, Methodology, Experiments, and Conclusion, although the order may slightly vary depending on the venue where they’re published. Are there sections that we could just skip? Yes. More specifically, reading them all, especially in the order as written, might even make you feel too bored to finish it up.

For explanation purpose, I first assume you are a beginning reader, and I take as an example a Google Deep Mind’s paper: “Playing Atari with Deep Reinforcement Learning”. This was the paper in 2013 when ‘Deep Learning’ had just started attracting many attentions in research community, but even now, reinforcement learning lovers would enjoy referencing it for their projects because of its scientific impact and clarity of description. The paper is on arXiv so anyone can download it for free.

Focus on ‘Head’ and ‘Tail’!

You as a beginner very likely haven’t studied ‘Deep Reinforcement Learning’ before. I’d recommend that you read Title, Abstract, and Introduction (Head) in order but skip the body part, such as Related Work and Methodology, to jump right into the tail sections, i.e., Experiments and Conclusion (Tail). Moreover, if you feel less interested while reading, you can even skip Introduction and Conclusion as well. That is, two possible methods could be:

Method 1) Title → Abstract → Introduction → Experiments → Conclusion

Method 2) Title → Abstract → Experiments

You might think that if you didn’t read the Methodology section, it’d be difficult to understand Experiments afterwards. However, if you tried any of the methods suggested above, you’d be surprised to realise that it’s quite fine not to read the bodies in comprehending what’s following. Let’s practice with the real paper then.

First of all, from the title “Playing Atari with Deep Reinforcement Learning”, you might feel uncomfortable with the term, Deep Reinforcement Learning. Actually, research papers are not as kind as college textbooks, because not all technical words’ definitions are included, and you’re basically assumed to be a non-beginner who has already been familiar with the terminology in the field. Therefore, you can choose to use an external resource to look up brief definitions, such as Wikipedia, if you’re already scared about what would go on:

Deep reinforcement learning (Deep RL) is a subfield of machine learning that combines reinforcement learning (RL) and deep learning …

Such a description on Wikipedia can also be tough, if you try to read all from top to bottom, so I would suggest only looking at first 1–2 sentences to only get main ideas. In this example, you could learn ‘Deep RL’ is a sort of a combination of reinforcement learning and deep learning, and it’d just be sufficient enough to turn back to the Deep RL paper.

Now, look at the abstract:

… The model is a convolutional neural network, trained with a variant of Q-learning, whose input is raw pixels and whose output is a value function estimating future rewards. We apply our method to seven Atari 2600 games …

As you can see, the abstract already provides a conceptual explanation of their proposed methodology with important keywords: ‘Convolutional Neural Network’ with inputs of ‘raw pixels’ and outputs of ‘estimated rewards’. In addition, the game ‘Atari 2600’ seems to be used for application of their methods. Now then, we go straight to the Experiment section!

Look at Figures and Tables!

In Experiments, you can also not read sentences of technical details you couldn’t clearly understand now as a beginner. Don’t feel guilty! Look at the visualised images and tables instead to find 1) what numbers there would imply and 2) how evaluations are performed in the field (of deep reinforcement learning).

These plots are put in Figure 2, in which each plot seems to show an increasing trend of line. You might feel an instant frustration since you don’t know what ‘Epoch’ in x-axis means. But the caption of the figure kindly explains:

Figure 2: The two plots on the left show average reward per episode on Breakout and Seaquest respectively during training …

Because the description says the line is the reward during training, you could naturally learn that higher ‘rewards’ are obtained as the training proceeds.

More details are also in a paragraph like this:

The leftmost two plots in figure 2 show how the average total reward evolves during training on the games Seaquest and Breakout. Both averaged reward plots are …

Below is the screenshot from Table1.

Even though you don’t know what Random, Sarsa, and Contingency are, it’d still be clear that the proposed model, Deep Q-Networks (DQN), leads the highest numbers (in bold) in most cases. The caption also says that the reported number indicates ‘reward’, as in Figure 2.

Moreover, you can find another paragraph offering more interpretations of this table:

The first five rows of table 1 show the per-game average scores on all games. Our approach (labeled DQN) outperforms the other learning methods by a substantial margin …

As expected, there are more details here, which seem aligned with our observations that the numbers are basically the scores from Atari games, and DQN is superior to others.

Done? Yes!

Finally, look at Conclusion!

This paper introduced a new deep learning model for reinforcement learning … Our approach gave state-of-the-art results in six of the seven games it was tested on …

This section basically summarises what we’ve seen so far. “Raw pixel”-based model in Abstract is just paraphrased as “a new deep learning model”, and the observations from Table 1 are as “state-of-the-art results” here. That is, we can ensure that our reading strategy has quite well covered what the authors wanted to deliver throughout the paper, although a large portion was actually ignored.

How about the aforementioned goal of your reading as a beginning reader? — ‘‘get a rough idea about the field and what the authors have done!’ Well, while looking through only selective contents, you’ve actually learnt the following:

1. Video games like Atari2600 have been used for validation and evaluation.
2. Reinforcement learning approach using pixel-based inputs has been proposed.
3. Increasing reward, or game score, during training is highly valued.
4. DQN has significantly outperformed other baselines.

It’d be surprising that as a beginner, you needed relatively short time to get all this knowledge. If you want to know more, you could spend more time with the paper, but for achieving the initial goal, this level of understanding could just be sufficient.

You might still feel guilty of not completing every single word from the paper, but I strongly believe that if you utilise your time to read other new papers instead of digging more into this DQN paper, you could more quickly gain an overall view of the field of reinforcement learning. Once after you’ve obtained more knowledge, you could come back to this paper to answer new questions popping into your head.

I hope this post has been insightful for you. I will then see you in next posts where I will also share other reading strategies for different goals of reading. Thank you!