1. Introduction

In 2021, the global market size for artificial intelligence had reached USD 328.34 billion (Fortune Business Insights, 2022). The impact of COVID- 19 and industry 4.0 reveals the importance  of artificial  intelligence  across  all  industries.  Manufacturing,  hospitality,  and tourism companies strongly demand artificial intelligence technology to lead their industries (Parvez, 2021; Rai et.al., 2021). With an expected growth rate of 20. 1% per annum, it is estimated that the artificial intelligence market size will expand to USD 1,394.50 billion in 2029 (Fortune Business Insights, 2022).

Further  expanding  the  market's  growth  requires  the  artificial  intelligence  community participants, for example, researchers and engineers, to deliver more research on artificial intelligence. Therefore, identifying top participants in the community becomes critical because it helps enlarge the community network and strengthen the quality of research by encouraging collaborations between top participants. In this report, the International Conference on Learning Representations (ICLR) is selected to analyze the contribution of the top participants in the community. ICLR is one of the top publications for research in deep learning (ICLR, 2022). According to Google Scholar (n.d.), it has the most paper published between 2016-2020 in the artificial intelligence category. Hence, analysis of ICLR can genuinely represent the overview of the artificial intelligence community.

2. Key Question

This report analyses the critical question: who is the most productive author from ICLR 2017-2021 and how is his/her performance? This question aims to identify past and present top contributors to the artificial intelligence community and its trend in the future. Not only will research on the productivity of authors, but the question will also justify the quality of prolific authors'  papers  and  the  most  productive  author's  submission  topics  and  collaboration performance. The question intends to enhance the quality of the paper and the chance of acceptance in ICLR by helping authors identify the best co-author to collaborate with and expand their authorship network. Also, connect authors with the most productive author who shares similar research. Identifying top authors in the community also builds up the reputation of top authors, allowing more peer reviews of their research. As a result, it may generate new insight with more discussion around their research and amplify the growth of the artificial intelligence market.

Although top authors contribute the most to the artificial intelligence community, they may not have the most contribution to each paper. This report also addresses a sub-question: are top authors mainly the first author of their paper? If not, who is the most productive first author from ICLR 2017-2021? The first author is the one who has the most significant contribution to the paper (Elsevier, n.d.). Thus, the first author's quality would significantly impact paper’s quality and acceptance probability. The sub-question hypotheses is that top authors are rarely the first author of their research because the first author requires a significant contribution to the work, which is impossible when top authors publish numerous papers every

year.

For the same reason, this report assumes that most first authors are fresh researchers, for example, graduate students and PhD candidates. Compared with top authors that are more known in the community, fresh participants required more reputation and support for their work to be seen by the community. This sub-question intends to introduce some of the top new researchers and their papers to the community and help them expand their authorship network.

3. Analysis Method

This report follows descriptive research (Cote, 2021; Dulock, 1993) and the analysis method is divided into four steps. First, the raw data is acquired from the online server. Then, the raw data is transformed into new data suitable for the analysis. Next, data visualization presents the research result. Finally, the analysis would generate insights into the past and current patterns and future trends of the top authors in the artificial intelligence community.

The  analysis  is  based  on  datasets  collected  using  OpenReview’s  API.  For  each  ICLR submission from 2017 to 2021, the submission ID, title, author list, author emails, author IDs, author affiliation, abstract, keywords, final decision, and peer review ratings are acquired (given files). In total, 7,615 submissions are collected: 477 in ICLR 2017, 911 in ICLR 2018, 1,419 in ICLR 2019, 2,213 in  ICLR 2020,  and 2,595 in ICLR 2021.  The datasets  acquired  from OpenReview’s API (given files) can address the key question; thus, no additional data is required for the analysis. However, data transformation must be performed to draw useful information for the analysis. In this report, data transformation is performed using Python.

To determine the performance of a submission, the average peer review rating is calculated for every  submission.  Submissions’  final  decisions  are  grouped  into ‘accept’  and ‘reject  or withdrawn’, where ‘reject or withdrawn’ includes submissions without final decisions. These papers  are  assumed  to be  withdrawn prior  to  the  final  decision.  In  ICLR  2017,  several submissions have a decision: ‘invite to workshop track’, and they are considered rejections.

Another critical assumption in this report is that there is no duplicate author name. In other words, the author’s name is used as a unique identifier. The reason for this assumption is outlined in section 7. For each author, total submissions and accepted papers in each conference are counted to determine the author’s productivity and performance. Also, the average rating of each submission for the author is grouped into a list. Then, the acceptance rate and average submission rating per author (average of papers’ average ratings) are calculated. The process is repeated for ICLR 2017-2021 and merged into one dataset. The cumulated acceptance rate and cumulated average submission rating per author from ICLR 2017-2021 are also calculated in the dataset.

To analyze first authors’ productivity, the first author’s name in each paper’s author list is defined as the first author because the contributions to the work determine the sequence of authors  (Bhattacharya,  2010).  Then,  the  same  data  transformation  process  as  above  is performed. The total number of records is less than the dataset above because there is only one first author in one submission.

After identifying the most prolific author, the keywords (topics) and co-authors’ frequency of the most prolific author’s submissions from ICLR 2017-2021 are counted from the raw datasets. The keyword and co-author frequency datasets are used to determine the author’s research topics and network. In addition, statistics of submissions with co-authors, such as acceptance rate and average submission rating, are calculated. They are used to determine the performance of each collaboration.

To begin the analysis, a multiple line chart illustrates the average submission of top authors compared with the average submission of all authors from ICLR 2017-2021. This chart aims to provide a general overview of how the productivity of top authors changes from ICLR 2017-

2021.

Next, bar charts and line charts are used to present the statistic of the productive authors, such as the total submission and average submission rating per author from ICLR 2017-2021. Bar charts present basic visualization of authors’ submissions and accepted papers, and line charts present  the  average  submission  rating.  They  provide  information  about  top  authors’ productivity and performance over the five conferences. Besides, multiple line charts illustrate the timeline of total submission, acceptance rate, and average submission rating of top authors from ICLR 2017-2021. It is used to present the change in top authors’ productivity and performance from ICLR 2017-2021

After distinguishing the most productive author, a violin chart illustrates the paper’s rating distribution in each conference. The violin chart aims to reveal any significant paper quality dispersion in each conference. Next, a word cloud visualizes the popular topics in the most prolific author’s papers and allows audiences to understand the major research areas of the most productive author. Also, a chord diagram presents the co-author network of the most productive author. The chord diagram also shows any collaboration between top authors.

Following the analysis of the critical question, bar charts and line charts are used again to present the productivity and performance of the top first authors. They also determine whether top authors have high submissions as the first author. However, multiple line charts are not used to show the submission timeline for the top first authors from ICLR 2017-2021 because their total submissions as first authors are all less than ten, and most of the first authors do not publish in all five conferences. Hence, presenting the timeline of the top first authors’ submissions will not provide sufficient information for the analysis.

4. Data Description

There are ten datasets acquired from the API of OpenReview. 5 of them (paperlist files) include the submission details in the five conferences. There are 477 records (submissions) in ICLR 2017, 911 records in ICLR 2018, 1,419 records in ICLR 2019, 2,213 records in ICLR 2020, and 2,595 records in ICLR 2021. The datasets consist of 10 columns: 2 columns of submission ID, title, author list, authors’ emails, abstract, keywords, summary, final decision, and peer review ratings. Peer review ratings range from 1 to 10, and there are multiple reviews for each submission. In ICLR 2017-2020, all submission summaries are missing, and several are missing in ICLR 2021, but they do not impact the analysis. In ICLR 2017 and 2020, several papers omit both the final decision and peer review rating. This report assumes that authors withdraw these submissions  before  peer  review  is  carried  out.  Hence,  they  are  classified  as ‘reject  or withdrawn’ in the analysis. In ICLR 2019 and 2021, one submission misses the peer review rating  in  each  conference.  It  may  be  due  to  errors  during  web  crawling;  hence,  these submissions are not included in calculating authors’ average submission rating, but they are regarded as valid submissions. In ICLR 2017, two submissions had invalid final decisions and peer review ratings caused by web crawling errors. Hence, manual correction is performed by editing the two submissions’ final decisions and peer review ratings using the paper information obtained from the OpenReview website.

Another five datasets (affiliation files) include the author’s details in the five conferences. There are 1,289 records (authors) in ICLR 2017, 2,748 records in ICLR 2018, 4,674 records in ICLR 2019, 7,682 records in ICLR 2020, and 9,673 records in ICLR 2021. In each dataset, records can be duplicated if the author has multiple submissions in the conference. Therefore, records in the dataset do not represent the number of authors participating in the conference. The datasets consist of 4 columns: author ID, name, affiliation, and email list. Nevertheless, these five datasets are not used in this analysis, given that they help identify the affiliation and country of each author, which is not the focus of this report.

After the data transformation process mentioned in section 3, five datasets are created to analyze the  critical  question  and  its  sub-question.  The  author’s  dataset  contains  16,591  authors’ information from ICLR 2017-2021. It has 31 columns: one column of author name, five columns of total submission, accepted paper, a list of papers’ ratings, average submission rating per author, the acceptance rate for the five conferences and the cumulated number from ICLR 2017-2021 (1+5*6). However, one record has a missing value in the ‘author’ column, which may occur due to data errors in original datasets. Hence, the record is removed from the analysis.

First author’s dataset contains 5,950 authors’ information from ICLR 2017-2021. It has 26 columns, and they are the same as the columns of the author’s dataset, except it excludes the list of papers’ average ratings for the five conferences and the total of the five conferences. Similar to the author’s dataset, there is one record with a missing value in the ‘author’ column, which is removed from the dataset.

Two of the remaining datasets are used for the most productive author’s chord diagram and word cloud. One of the datasets contains the collaboration frequency of 230 co-authors that worked with the most productive author, and the other datasets contain the frequency of 173 topics for the most productive author’s paper.

Finally, there is a dataset containing the submission information of 230 co-authors that worked with the top authors. It has eight columns with information in authors’ names, lists of average rating per paper, average submission rating, a maximum and a minimum submission rating, the number of collaborations, accepted papers, and acceptance rate.

To  begin  the  analysis,  some  information  from  the  raw  datasets  about  the  conference  is visualized.

Figure 1: (left) Total Submissionsfrom ICLR 2017-2021.

(right) Average Submission Ratingfrom ICLR 2017-2021.

Figure 1 shows that the total submission in ICLR has increased from around 500 in ICLR 2017 to over 2,500 in ICLR 2021. The number of accepted papers has also increased from ICLR 2017-2021. However, there is a decreasing trend in the acceptance rate from ICLR 2017-2021. There is over 40% of acceptance rate in ICLR 2017, but the acceptance rate has dropped by almost 10% in 5 years. The figure also indicates a positive relationship between acceptance rate and average submission rating per year. In ICLR 2020, on average, there is a substantial decrease in the average submission rating and acceptance rate. It is potentially caused by tighter peer review or a decline in submissions’ quality. Further investigation of the ICLR 2020 peer review can help understand the drop in paper ratings and acceptance rate.

5. Results Presentation

To start off, a multiple line chart is used to show how the productivity of top authors changes compared to all authors from ICLR 2017-2021.

Figure 2 shows that the submissions of the top 10 authors have doubled in five years, from an average of 7.6 submissions in ICLR 2017 to 16.7 submissions in ICLR 2021. It indicates that the current ten most productive authors published twice the number of papers five years ago, and there is an increasing trend of total submissions for prolific authors. In addition, the average submissions  for the top  100  and  1,000  authors  are increasing too, but  at  a  slower pace. Nonetheless, the average submission of all authors seems to have a constant trend. Given that the total submission increases every year (figure 1), the constant trend indicates that there are more authors participating in the conference and many authors with only one submission in

ICLR 2021.

To further investigate the top prolific authors from ICLR 2017-2021, bar charts and line charts illustrate their cumulated submissions, acceptance rate, and average rating.

Figure 3: (left) Total Submissions of Top 10 Authorsfrom ICLR 2017-2021.

(right) Average Submission Rating of Top 10 Authorsfrom ICLR 2017-2021.

From figure 3, Sergey Levine is the most productive author in ICLR 2017-2021, with over 100 total submissions in 5 years and a 52.8302% acceptance rate. Among the top 10 authors, Sergey Levine has the highest acceptance rate and second-highest average submission rating. Indicates that Sergey Levine can maintain the paper quality on average while publishing more than 20 papers on average per conference. Besides Sergey Levine, Yoshua Bengio and Pieter Abbeel also have more than 60 submissions in five conferences. However, their acceptance rates and average submission ratings are lower than Sergey Levine's. It means that Yoshua Bengio and Pieter Abbeel are authors with significant productivity, but the quality of their work is not ensured. Especially for Pieter Abbeel, who has the second-lowest acceptance rate and average submission rating in the top 10 authors.

The remaining seven authors have similar submissions from ICLR 2017-2021. Richard Socher is the only author with a more than 50% acceptance rate. He also has the second-highest acceptance rate in the top 10 authors. Ruslan Salakhutdinov has the highest average submission rating in the top 10 authors, with an almost 50% acceptance rate. Indicates that Richard Socher and Ruslan Salakhutdinov can consistently deliver high-quality papers, although they have less than half the productivity of Sergey Levine.

The result shows that the top 5 authors have either or both high productivity and high paper quality.  The  following  multiple  line  charts  illustrate  the  trend  of the  top  5  authors  in submissions, acceptance rate, and average submission rating over ICLR 2017-2021.

Figure 4: Multiple Line Charts of Top 5 Authorsfrom ICLR 2017-2021 on Total Submissions (left),

Acceptance Rate (middle) and Average Submission Rating (right).

Figure 4 shows that Sergey Levine’s submission tripled from 2017 to 2020. Although there is a slight decrease in submission in ICLR 2021, he is the most productive author in recent ICLR across the top 5 authors. He maintains an over 50% acceptance rate except for ICLR 2020. Given that the overall acceptance rate and average paper rating have dropped in ICLR 2020 (see figure 1), he is the only top 5 author who preserves a 40% acceptance rate in that year. In addition, He has the highest average paper rating among the top five authors in ICLR 2017, 2019, 2020, and 2021. Indicates that he has outstanding paper quality among prolific authors

on average.

Yoshua Bengio has had a decreasing trend in publication since ICLR 2019. Among the top 5 authors, despite having the lowest acceptance rate from ICLR 2017-2019, he has the highest acceptance rate with an average paper rating of over 6 in ICLR 2021. The improvement in paper quality may be due to the reduction in submission, where he can concentrate on less research and enhance the work undertaken. Compared with Yoshua Bengio, Sergey Levine has over ten more submissions than Yoshua Bengio in ICLR 2020 and 2021 and still maintains a similar average paper rating as Yoshua Bengio.  Showing  Sergey Levine’s ability to pursue high productivity and quality.

Pieter Abbeel and Caiming Xiong have an increasing trend of submissions from ICLR 2017- 2021. However, the quality of their work deteriorates during the period. They record the worst acceptance rate and average rating per paper from ICLR 2020-2021.

Richard Socher has the minor submissions in the top 5 authors. He has a high quality of work because he can maintain the second-highest acceptance rate in his conferences. However, his submissions in ICLR 2021 dropped from 11 to 5.

The result shows that Sergey Levine is the most productive author and has strong paper quality on average. Next, a violin chart is used to present the distribution of his work in the past conferences. It is utilized to observe the quality of each paper he published in ICLR.

Figure 5: Violin Chart of Sergey Levine’s Submission Ratingfrom ICLR 2017-2021.

The black box inside the violin chart indicates the inter-quartile range of the papers’ ratings, and the white dot indicates the median of the papers’ ratings.

From figure 5, the overall performance for Sergey Levine is the highest in ICLR 2017, with over 50% of submissions scores 7 (good paper, accept) or higher. However, it does not indicate that the performance in ICLR 2017 is the highest because Sergey Levine only has seven submissions in ICLR 2017 (see figure 4). The average rating per paper in ICLR 2017 is the highest from ICLR 2017-2021 (see figure 3).

Except  for  ICLR 2020, half of the  submission  each year reaches  an  average  score  of 6 (marginally above the acceptance threshold). Indicates that more than half of his submissions in  ICLR  2017,  2018,  and  2021 meet  the  criteria  of being  accepted  into  the  conference. Moreover, over 75% of submissions in each conference are above an average rating of 5

(marginally below the acceptance threshold), except for ICLR 2020. Indicating that his paper quality is consistent, and most of his paper ensures high quality of study and research.

Nonetheless, the ICLR 2020 submission rating distribution is the most dispersed. There are more counts at the bottom of the violin chart, with more than 25% of submissions receiving an average rating of 4 (ok but not enough - rejection) and only 25% of submissions above an average score of 6 (marginally above the acceptance threshold). It is the only conference he records a paper’s average rating below 3 (clear rejection). Although the overall rating in ICLR 2020 declines dramatically, Sergey Levine obtains the best performance on average among the top 5 authors. It is uncertain whether the decline in papers’ ratings is caused by tighter peer review or decreased quality of submissions. Hence, further research is required to understand what happened to Sergey Levine’s paper rating in ICLR 2020.

The  above  result  illustrates  that  Sergey  Levine  is  the  most  productive  author  and  has consistently delivered high-quality papers in the past five conferences. Next, a word cloud is presented to show the topics of Sergey Levine’s research. The larger the font size of the topic, the more frequent the topic appears in Sergey Levine’s papers.

Figure 6: Word Cloud of Sergey Levine’s Research Topics

(note: one paper is not limited to one topic).

Sergey  Levine’s most  frequent  research  topic  is  Reinforcement Learning.  Deep Learning, Meta-Learning, and Imitation Learning are popular topics in Sergey Levine’s research. A lollipop chart presenting the top 50 research topics of Sergey Levine can be found in Appendix I.

The majority of Sergey Levine’s papers focus on Reinforcement Learning. Around 55 papers,  50%  of  submissions  published  in  ICLR  2017-2021,  cover  Reinforcement Learning (see appendix I). While the remaining research topics have less than 20 frequency, Deep Learning, Meta-Learning, and Imitation Learning appear more than ten times. Indicates that at least 10% of Sergey Levine’s research topics focus on each of the three topics.

Lastly, this report analyzes Sergey Levine’s co-author network to identify any connection with prolific authors and help the audience to connect with Sergey Levine through his network. Thus, a chord diagram presents the most frequent co-authors Sergey Levine has collaborated with from ICLR 2017-2021.

Figure 7 presents the co-authors collaborating more than four times with Sergey Levine. In total, there are 18 of them. Among the top 18 collaborators, there are three top prolific authors: Yoshua Bengio, Pieter Abbeel, and Chelsea Finn. Indicates that there is connect between Sergey Levine and top authors. One of the top prolific authors (see figure 3), Chelsea Finn has the most collaboration frequency with Sergey Levine. This was followed by Abhishek Gupta, Pieter Abbeel, the author with the third-highest submissions in ICLR 2017-2021, and Benjamin Eysenbach. Given that Sergey Levine has more than 100 submissions in ICLR 2017-2021, figure 7 indicates that many of Sergey Levine’s co-authors only have infrequent collaboration (less than five times).

Figure 7: Chord Diagram of Sergey Levine’s Co-author Network.

From table 1, Sergey Levine’s collaboration with Benjamin Eysenbach produces the highest average submission rating and acceptance rate among the top 5 frequent co-authors. Followed by Chelsea Finn, who delivers the second-highest average submission rating and acceptance rate.  Collaboration with Abhishek  Gupta  delivers the  lowest  acceptance rate  and  average submission rating. The result may indicate that Sergey Levine will deliver more quality research when working with Benjamin Eysenbach and Chelsea Finn. Nevertheless, the research quality may decline when working with Abhishek Gupta.

However, the acceptance rate and average submission rating may not truly represent the quality ofresearch. From section 4, the acceptance rate and average submission rating declined in ICLR 2020. Hence, the statistic in table 1 is dependent on which year the authors worked together. For instance, if Sergey Levine teamed up with Benjamin Eysenbach in ICLR 2020, their paper would have a higher chance of receiving a low rating and not being accepted. Hence, further research is required to determine how Sergey Levine’s co-author network affects the research quality.

Lastly, this report addresses the sub-question outlined in section 2. Of the top 5 prolific authors, only Yoshua Bengio and Caiming Xiong have published as the first author from ICLR 2017- 2021. Yoshua Bengio had one submission as the first author in ICLR 2020 with an average rating of 6.3333 and one in ICLR 2021 with an average rating of 7.6667. The conference accepts both submissions. Caiming Xiong had one submission as the first author in ICLR 2017 with an average rating of 8 and one in ICLR 2018 with an average rating of 7. Both submissions are also accepted. Therefore, the result concludes that top prolific authors are not mainly the first author. It proves the first part of the hypothesis in the sub-question that the top authors should rarely be the paper’s first author, given the significant contribution required by the first author.

To illustrate the top prolific first author and their research quality, bar chart and line chart are used to illustrate their cumulated submissions, acceptance rate, and an average submission rating.

Figure 8: (left) Total Submissions of Top 10 First Authorsfrom ICLR 2017-2021.

(right) Average Submission Rating of Top 10 First Authorsfrom ICLR 2017-2021.

From figure 8, the number of papers published by the same first author from ICLR 2017-2021 is less than ten. There are three authors with eight submissions as the first author: Anirudh Goyal and Sanjeev Arora have six accepted submissions, and Xinyun Chen has four accepted submissions. Even though two authors have the same acceptance rate, Sanjeev Arora has the highest average submission rating for submissions as the first author within the top 10 first authors. Hence, the result concludes that Sanjeev Arora is the most productive first author with the highest paper quality on average.

Furthermore, there are four authors with seven submissions as the first author. Dan Hendricks is the only author with over 50% acceptance rate and an average submission rating of 6. There are also multiple authors with six submissions as the first author but not covered in figure 8. However, only Yu Bai has an over 50% acceptance rate. It indicates that the paper quality of the prolific first authors varies. Half of the top first authors have less than a 40% acceptance rate, compared with only two of the top 10 authors with less than a 40% acceptance rate (see figure 3).

Within the top 10 first authors, only Xinyun Chen has participated in all five conferences, and most first authors missed one to two conferences. For example, Sanjeev Arora, who has the highest submissions as a first author and average rating, has participated from ICLR 2017-2020.

From manual check, 5 of the top 10 first authors are senior researchers from IT companies or professors from universities. Hence, the hypothesis that most of the prolific first authors are fresh researchers in the artificial intelligence community is rejected. This report concludes that many senior participants, such as senior researchers and professors, will still frequently publish papers as the first author.

6. Insight

This report addresses the contribution of top authors and identifies the top first authors from ICLR 2017-2021. There has been an increasing trend of papers published by prolific authors in the past  five  conferences,  even  for the top  1,000  authors  (see  figure 2).  In  ICLR 2022, productive authors are expected to publish more papers. It is also expected to have more authors participate in the ICLR 2022, but many authors will have only one submission at the conference.

The result in section 5 illustrates that Sergey Levine is the most productive author in ICLR 2017-2021. He is the only author with over  100 submissions from ICLR 2017-2021 and achieves the highest acceptance rate among the top 10 authors. However, Sergey Levine has a decreasing trend of acceptance rate from 70% in ICLR 2017 to about 54% in ICLR 2021. The decrease in acceptance rate is potentially caused by a substantial increase in submission every year or a tighter acceptance policy by ICLR since the acceptance rate of all submissions is decreasing (see figure 1). In general, Sergey Levine’s submissions receive high peer review ratings. Except for ICLR 2020, half of his submissions each year reach an average peer rating of 6. It indicates that at least half of his submissions meet the requirement of being accepted in the conference. In ICLR 2020, he experienced a significant rating disparity in his submissions. Only 25% of papers received an average rating of more than 6, and several papers had an average score of less than 3. It is reasonable to believe that there is an overall poor performance in ICLR 2020 because ICLR 2020 had the lowest acceptance rate and lowest mean rating in ICLR  2017-2021  (see  figure  1). Nevertheless,  Sergey  Levine  still  maintains  the  highest acceptance rate and average submission rating among the top 5 authors in ICLR 2020.

Therefore,  Sergey Levine demonstrates the most potent productivity and high submission quality from ICLR 2017-2021. However, his submissions drop slightly in ICLR 2021. Thus, Sergey Levine is expected to publish fewer papers in ICLR 2022 but remain the top productive author in ICLR 2022 and continuously deliver outstanding papers.

Sergey   Levine’s   research   topic   focuses   on   Reinforcement  Learning,   Deep Learning, Meta-Learning, and Imitation Learning. Sergey Levine has over 50% of   ICLR   2017-2021   submissions   covering   Reinforcement  Learning.   Deep Learning, Meta-Learning, and Imitation Learning each cover around 10% of Sergey Levine’s submissions. Hence, authors with similar research areas as Sergey Levine are encouraged to collaborate with Sergey Levine. Given that majority of Sergey Levine’s papers focus on Reinforcement Learning, it is expected that Sergey Levine will continuously concentrate on Reinforcement Learning studies in ICLR 2022.

Sergey Levine’s co-author network exhibits the network between prolific authors. He has frequently collaborated with two of the top 10 authors. Almost 30% of Sergey Levine’s papers are written with Chelsea Finn, who has the ninth-highest submissions from ICLR 2017-2021. Collaborations with Chelsea Finn also demonstrate high-quality research on average. Indicates Sergey Levine has a strong co-author network with Chelsea Finn, and their collaboration is expected to continue in ICLR 2022. Hence, authors that wish to connect with Sergey Levine can try connecting through Chelsea Finn or other co-authors (see figure 7). Besides, Sergey Levine’s collaborations with Benjamin Eysenbach deliver the highest performance among top co-authors that worked with Sergey Levine. However, the result does not consider which conference they worked on together. Hence, further investigation is required to determine Sergey Levine’s performance with his co-author network.

In addition, Richard Socher and Ruslan Salakhutdinov also deliver high-quality papers within the top 10 authors despite not being as productive as Sergey Levine. Of the top 10 authors, Richard Socher has the second-highest acceptance rate, while Ruslan Salakhutdinov has the highest average submission rating. However, Richard Socher’s submission dropped over 50% from  ICLR  2020-2021.  Thus,  this  report  expects  Richard  Socher  to  have  around  five submissions in ICLR 2022.

Top prolific authors are rarely the first author of their submissions, which indicates that they do not have the most contribution to their submissions. Only Yoshua Bengio and Caiming Xiong in the top 5 authors have submissions as the first author. Each of them published two papers as the first author in the past five conferences. Although most prolific authors do not usually publish a paper as the first author, the quality of their work is high because the conference accepts all four submissions. The result indicates that Sergey Levine will not publish a paper as the first author in ICLR 2022 and further encourages first authors to collaborate with Sergey Levine, Richard Socher, and Ruslan Salakhutdinov, the top  10 authors that deliver highquality papers on average.

Anirudh Goyal, Sanjeev Arora, and Xinyun Chen are the most productive first author from ICLR 2017-2021, with eight submissions as the first author. Anirudh Goyal and Sanjeev Arora have the highest acceptance rate of 75%. Among the most productive first authors, Sanjeev Arora has the highest average submission rating as the first author. Indicates that Sanjeev Arora is the highest productivity first author with the highest paper quality on average. However, most prolific first authors do not participate consecutively from ICLR 2017-2021. Sanjeev Arora did not participate in ICLR 2021, and it is uncertain whether he will participate in ICLR 2022.

Unlike prolific authors (most acceptance rates are around 50%, and most average submission ratings are above 5.3), the qualities of prolific first authors’ submissions vary. The acceptance rate ranges from 0% to 75%, and the average submission rating range from around 4.6 to 6.3. However, a limited conclusion about the first authors’ quality can be drawn from the analysis. The total papers published by the same first author in five conferences are all less than ten.

This report highlights some of the top first authors that perform well from ICLR 2017-2021. Anirudh Goyal, Sanjeev Arora, Xinyun Chen, Dan Hendrycks, and Yu Bai have published more than five submissions in five conferences with at least a 50% acceptance rate.