The 200 deep learning papers I read in 2019

The 200 deep learning papers I read in 2019

Half a year ago, I wrote my first post on Medium about my journey of routine paper reading, as part of the fulfillment of my New Year’s Resolution. At that time, I concluded that this daily activity of paper-reading is crucial to keep my mind active and abreast of the latest advancement in the field of deep learning. Now on the eve of the new year of 2020, I can proudly say that I executed my 2019 new year resolution of “reading at least one new paper per week” with flying colors. In the second half of 2019, I read 116 papers and kept organized notes. If you are interested, the structured notes of these papers are listed in this following Github repo.

What did I read in 2019?

Together with the 76 papers in the first half of the year, the total number of papers I read in 2019 is 192 papers (including the papers I’ve skimmed during writing notes this number should be a bit more than 200). That is pretty close to the stretch goal of one paper per workday . Of course, I slacked off sometimes or was not disciplined to wake up early enough to finish one paper, but I try to make that up during the weekend. To gain a better understanding of my reading habits, I did a quick visualization, inspired by the ICLR OpenReview visualization . The code for generating the graphs, including the titles of the papers, are here .

About 140 of the 200 papers are peer-reviewed publications and the rest are ArXiv preprints. By publication year, about half of the peer-reviewed papers are from within this year of 2019, with the earliest one dating back to 2012. (Just in case you wonder, the only 2012 publication is the original KITTI dataset paper from CVPR 2012.)

As of publication venue, one-third of them are from CVPR, and together with ICCV, they make up about half of the papers I read. I find CVPR papers to be more application orientated, and NIPS and ICLR papers are a bit too theoretical to my taste and for the field of autonomous driving that I am focusing on. Interestingly, I constantly find many hidden gems in workshop papers at NeurIPS/ICLR.

In the latter half of 2019, I have been really disciplined to stick with my stretch goals, except in December. So the task of reading one paper per workday is doable, but perhaps with a caveat (see below).

Things that worked well for me

• Keep on waking up early and reading papers in the morning. This is a good wake-up exercise and prepares me better for the day.
• Read papers by topics. This helps avoid too frequent context switching and keeps good continuity in reading papers in a related field. Also, this motivates me to write mini-review blogs on the topics. In the second half of 2019, I have written a handful of review blogs on Medium.
• When writing blogs, it is better to start small, focusing on one small area of a much broader topic. When I realized that I have read quite a few papers on monocular 3D object detection by September, I decided to write a review about it, at least for my record. However, this is a daunting task as there have been so many papers recently published in 2019 (actually the majority of the paper on this topic were in retrospect). So I started small by writing one post about the multi-bin loss method used to regress the orientation of objects ( Multi-modal Target Regression ). Then I realized that the terminology of yaw orientation in literature is quite confusing so I started another post that is perhaps the least technical but mainly conceptual ( Orientation Estimation in Monocular 3D Object Detection ). After these two posts, I feel more comfortable expanding the topic of my next post to a particular approach of the study that lifts 2D to 3D in postprocessing ( Lifting 2D object detection to 3D in autonomous driving ). Finally, when I feel I have enough grasp of the trend in this field, I wrote an overarching review summarizing different approaches to solve the seemingly ill-posed monocular 3D object detection problem towards the end of November ( Monocular 3D Object Detection in Autonomous Driving — A Review ).

Things that need improvement

• I still tend to get bogged down in details. I tend to dive directly into papers in a linear manner with the same amount of attention to detail. Professor Srinivasan Keshav from Cambridge University has written a short article on how to read a scientific paper in three passes (Summary in Chinese can be found on 知乎 ). MIT also has a guideline on reading paper under time constraints . I should definitely start practicing these more disciplined methods to guide me to decide when to skim through and when to dive deep. Reading one paper a day in full detail is not sustainable for me as a software engineer — although this may be doable and even make sense to do so for full-time researchers.

• Quality over quantity. Read high-quality papers, and spend more time with higher quality papers. Low-quality papers often have conflicting or inconclusive results (maybe due to premature and insufficient experiments). This is particularly hard to foster a systematic and coherent understanding of a particular topic. How to decern these papers? First, I should have a watch list of authors and research groups known to produce high-quality papers. Second, I should skim through the highly scored or best paper nominees of major conferences. Third, read slightly older but seminal papers (one or two years old) with many citations . Fourth, find papers with released source code and read code.
• Talk is cheap, show me the code. Spend more time reading source code on Github. It dawned on me many times that the code does not always reflect what the paper claims. My 2020 New Year’s Resolution should be, 1) reading and playing with one high-quality GitHub repo every month and write review notes about code-reading; 2) reading 10 papers in-depth with three passes a month (which takes a couple of hours each). This is roughly half of the speed I am plowing through the deep learning paper pile, and this should give me more time to digest.