FAQ for Working with Mihai
This document aims to answer common questions that incoming PhD/MS students have about working with Mihai.
The document is written by Mihai, so "I"/"me" means Mihai. (I thank Becky Sharp for feedback!)
New students should be signed up for the following mailing lists:
- clulab@list.arizona.edu - List for all people in the lab, including students, staff, and faculty.
- nlp-read@list.arizona.edu - List for people interested in our NLP reading group. This includes many people outside of the lab, the university, and even the country.
- clulab-students@list.arizona.edu - List of all students in the lab. This includes students advised by any of the faculty in the lab. No faculty or staff allowed.
- mihais-cs-students@list.arizona.edu - List of grad students advised by me. Initially, this included only CS students (hence the name), but has since grown to include students from other departments.
Students should have access to the following Google calendars:
- "CLU Lab" - calendar for CLU events that apply to the entire lab, e.g., our reading group.
- "Mihai's Individual Meetings" - calendar to schedule individual meetings with me (see the Scheduling Meetings section).
I am supposed to give you access to these mailing lists and calendars. If you think you are missing any of these, let me know.
Please see this Wiki page and this one for information on our lab's research servers and how to use them. Email me to get an account on the three research machines: amy, jenny, and clara. As an UA student, you also have access to the university's high-performance cluster (HPC). Please see https://public.confluence.arizona.edu/display/UAHPC/HPC+Documentation for more information.
The CLU lab has a GitHub account here: https://github.com/clulab. Ask me to give you read-write access to it. The philosophy of the faculty in the CLU lab is that all software should be open-source and shared between all people in the lab (and the wider NLP community). As such, we highly encourage you to use this organization for your software development. A few important repositories in this organization are:
- processors - Library for generic NLP utilities, e.g., part-of-speech tagging, parsing, semantic role labeling. This library also includes Odin, our rule-based framework.
- research - Repository for research ideas that may or may not work. Create a new folder in this repo for things you would like to try, but you are not sure they will amount to much. If they end up growing into serious projects, we usually create separate repos at that point.
- releases - Library of code releases associated with publications. I ask that a copy of the stable version of the software that you published be stored here, for easier access and reproducibility.
In general, I do not have a fixed weekly schedule for individual meetings. This is for two reasons: (a) my schedule changes too much from week to week, and (b) most students benefit from the flexibility. For this reason, it is the responsibility of the students to schedule their individual meetings. This can be done by booking a slot in the calendar named "Mihai's Individual Meetings" in the blocks also marked as "Mihai's Individual Meetings." These blocks also indicate if I am available in person, remotely, or both. I will add blocks for the coming week the weekend before.
For example, here is a block that has been divided into 4 individual meeting slots:
Please do not book a meetings outside of these blocks. If you can't find an available slot, email me, and I will find a time (usually within 24h).
Booking a meeting is a commitment. Please be there and be on time. You can of course cancel meetings, but please do it as much in advance as possible.
I prefer structuring meetings along Agile principles. That is, most meetings should follow this agenda:
- What did you accomplish since the last time we met?
- What is on your to-do list for next time?
- Is there anything blocking your progress?
- What can I do to help?
It is inevitable that research directions are criticized by your advisor during meetings. However, please keep in mind that this is not criticism on you. I (as well as the rest of the department) believe in you, and want you to succeed. However, ideas should be criticized for at least two reasons:
- The PhD is an environment where ideas can fail safely. Nevertheless, it is important to fail quickly, so you can hopefully converge on a successful idea that leads to publication quickly.
- I do not want you to be blindsided by insincere positive comments when you hit the job market. It is your advisor's job to make sure that you are prepared for real jobs, some of which do not have the safety net that the PhD offers.
It is likely that I have been around the block more than you so you should listen to my observations. However, it does not mean that I am always right. You are more than welcome to bring counter arguments to the discussion! Respectful disagreements very often lead to better ideas.
Natural language processing is heavily inter-disciplinary. It covers computer science, linguistics, math, engineering. Thus, it is practically impossible for anybody (myself included!) to know everything needed for NLP research. Please do not be afraid to tell me that you don't know or understand something! If I don't know what you don't know, there is no way for me to help you.
Some of the questions you might have are probably answered in our "Deep Learning for Natural Language Processing: A Gentle Introduction" book.
On this topic, I will first quote from a few smart people I respect and try to emulate.
Stephen Kobourov says:
"Take your research seriously:
- Classwork and homework are important, doing your TA/grader duties is important; however, nothing is more important to a PhD student than their research.
- You are responsible for driving your research projects forward, for reading relevant papers, following up your own ideas, and following up on ideas that come up during research meetings.
- You should have a clear idea about what work you need to do so that you work towards at least one publication at any time.
- When you agree to something verbally or over email, this is a commitment. For example, written weekly reports means that every week you will prepare weekly progress reports; a good paper draft by next week's meeting means that you are going to prepare a good paper draft by next week's meeting.
- During regular school times (Fall and Spring semesters) I expect my students to reply to emails within 24 hours."
Ronald Azuma identifies the following traits as most important for a successful PhD in his famous "So Long and Thanks for the Ph.D.!” essay: initiative, tenacity, flexibility, interpersonal skills, organizational skills, etc. Note that the first three are initiative, tenacity, and flexibility. Not your programming or math skills. All these are learnable if you have the initiative and tenacity. But the probability of completing your PhD goes down considerably without these fundamental traits.
Eugene Vinitsky list his expectations for his PhD students in this document. I generally agree with his points, with the exception that I do not expect first-semester students to commit 15--20 hours/week on research due to course and TAship/RAship overhead. (Unless your research is included in the RAship.)
To the above I would add that you have to take care of your mental health. Research is hard on all of us: there will be days when nothing works; you will receive paper rejections, etc. To manage all this, you have to learn how to disconnect from work. To this end:
- Try to find a hobby that is not related to your work;
- Go to the gym, or do other physical activity;
- Find a group of friends outside (or in addition!) to the people in the lab.
Lastly, do not compare your PhD against others (see below for details)! Everybody learns and progresses differently.
In general, my requirements for a PhD dissertation are that it should contain work equivalent to three publications in solid conferences. This should be work that you "own," i.e., you are the first author on these publications.
However, these general requirements deserve clarification:
- People organize their work differently: some students prefer to package their work in shorter and more focused conference papers. Some prefer the longer format offered by journal publications. For this reason, the work included in the dissertation may be organized in different ways for submission: three conference papers, two journal papers, one journal papers and several conference papers, etc.
- The work included in the dissertation should be submitted to good venues, but not necessarily accepted. We do not fully control the latter part of the process, so it is unfair to enforce it on students. As long as I agree that the work has high-enough quality for a good NLP venue, you are Ok. To understand how conferences are ranked, you can use the CORE Conference Portal. Similarly, the CORE Journal Portal provides rankings of computer science journals. In general, A and A* venues are great. B and C are just Ok.
- You have to drive the work that goes on in your dissertation. Typically, this means that you are the first author on the papers generated from this work. Having said that, I encourage collaborations, which normally lead to co-authorship on the respective publications. If you are not the first author on a publication, you can use the part you contributed in your dissertation, but cannot claim the whole paper.
PhD students are evaluated by the department at the end of each semester. The result of this process is either a Satisfactory (SAT) or Unsatisfactory (UNSAT) evaluation. Students who have UNSAT evaluations in two consecutive semesters have to leave the program.
In general, students are evaluated based on three criteria:
- Maintaining the minimum GPA required by the Graduate College;
- Performing the tasks that are part of their TA/RAship;
- Performing research that has the potential to become a dissertation.
In the first semesters, the weight assigned to the last criterion is lower because the overhead of starting a PhD is high, e.g., you have to take courses, you have to adjust to being a TA or RA, etc. However, I expect all students that work with me to engage with research that is potentially publishable from the first semester. Of course, the amount of work you invest in research will be smaller in the first semesters due to the other tasks you have to work on as a PhD student. But you need to commit some time to research every week. To extrapolate what Gichin Funakoshi said about karate: Research is "like boiling water; without constant heat, it eventually returns to its tepid state."
The importance assigned to the above criteria change starting with the fifth semester, when PhD students have to show the ability to drive research on their own (see also the department's requirements for the fifth semester portfolio).
To avoid subjectivity in these evaluations, the entire department votes on evaluation results for: (a) all intermediate milestones, i.e., 5th semester portfolio and comprehensive exams, and (b) any time an UNSAT is recommended by the graduate affairs committee (GAC). In particular, the following process is followed in a dedicated faculty meeting:
- The GAC recommends an evaluation result that takes into account feedback from the student's advisor and other faculty that interacted with the student;
- The recommendation is discussed with all faculty in the faculty meeting;
- The advisor leaves the room and more discussion happens between the remaining faculty; and
- Lastly, a vote is tallied between all faculty excluding the student's advisor.
Not everybody who starts a PhD will complete it. Deciding to stop your PhD is a major decision that is often very painful. Most students have committed considerable time to their PhD. In many cases, part of their identity becomes associated with doing a PhD (and planning for the corresponding future). So, I definitely understand that this decision is extremely difficult to make. However, if it gets to the stage where you, me, or both of us observe that is not working out, it is better to stop after k years than continuing for n years, where n >> k, and still not graduating.
Note that completing a PhD is not correlated with your intelligence or your value as a human being, but with the traits identified by Ronald Azuma as important, e.g., initiative, tenacity, flexibility (see "What Makes a Good PhD" above). Ultimately, these traits kick in only if you are passionate about what you are working on and research in general. This is a question that you have to answer for yourself.
Students who stop a PhD before completing it can still obtain a MS in CS (if you don't have one already). This is because the coursework for the PhD is almost identical with the MS coursework. Thus, a good time to re-evaluate your decision to commit 2+ years to completing a dissertation is after you complete the MS requirements.