This guide is inspired by my previous article, “Doing well in your courses,” which shared tips and tricks I learned during my undergraduate studies. That earlier piece received positive feedback, and in the same spirit, now that I’ve completed my PhD, I wanted to create a similar retrospective document, hoping it will be beneficial to others embarking on this challenging path. Unlike the undergraduate guide, crafting this one was considerably more complex. The PhD experience is far more varied and individualized. Therefore, many points are likely to be debated, and a significant portion will reflect my personal experiences, particularly within Computer Science, Machine Learning, and Computer Vision research. However, disclaimers aside, let’s delve into the essential advice.
Preliminaries: Is a PhD Right for You?
Image alt text: A group of PhD graduates in academic regalia, symbolizing the culmination of years of study and research, representing the achievement of a doctoral degree.
Before we dive into the survival rules for navigating a PhD, let’s first address a fundamental question: Should you even pursue a PhD? I was fortunate to have an early inclination towards doctoral studies. My reasons, however, were not entirely based on profound considerations. Firstly, I genuinely enjoyed learning and wanted to expand my knowledge as much as possible. Secondly, I aspired to emulate Gordon Freeman from the game Half-Life, a character with a PhD in theoretical physics from MIT. I was captivated by that game.
But what if you approach life decisions with more pragmatism? Is a PhD the right path for you? A helpful Quora thread offers valuable insights, and the following summary borrows and rephrases several points from Justin, Ben, and others in that discussion. I’ll assume the alternative you’re considering is employment at a medium to large company, a common career path for many. Ask yourself if the following aspects appeal to you:
Freedom. A PhD provides considerable freedom to explore and learn about topics that genuinely interest you. You are in the driver’s seat. While your advisor will provide guidance, you’ll generally have far more autonomy than in many other professional settings.
Ownership. The research you produce will be your own intellectual property. Your accomplishments will bear your name. In contrast, within a large company, it’s common to feel like a less distinct contributor. The feeling of becoming “a cog in the machine” is a common concern.
Exclusivity. Acceptance into top PhD programs is highly competitive. You’d be joining a select group of individuals, a few hundred in number, compared to the tens of thousands who join large companies annually.
Status. Whether justified or not, pursuing and obtaining a PhD is culturally recognized and respected as a significant achievement. Plus, you get to be called “Doctor,” which holds a certain prestige.
Personal Freedom. As a PhD student, you largely manage your own schedule. Want to sleep in? Need a day off for a short trip? Generally, that’s acceptable. What matters most is your research output. While some advisors are more structured than others, and some companies offer flexible work arrangements, the level of autonomy in a PhD is fundamentally different.
Maximizing Future Options. Embarking on a PhD doesn’t limit your future career choices. The path from PhD to various careers is well-trodden, but transitioning from industry to a PhD for an academic/research career is statistically less common. Furthermore, particularly in applied Machine Learning, PhD graduates and even dropouts are often highly sought after, potentially leading to more interesting roles and higher starting salaries. Maximizing options for your future self is a sound guiding principle.
Maximizing Variance in Experience. You’re likely young, and there’s no immediate rush. After a PhD, you could potentially spend the next 50 years in industry. Opting for a PhD allows for greater variety in your life experiences early on.
Personal Growth. A PhD is an intense period of rapid learning and personal development. You’ll gain deep knowledge and master self-management and resilience. Top PhD programs also offer a concentration of exceptionally bright peers who can become lifelong friends and collaborators.
Expertise. A PhD is likely your only opportunity to deeply immerse yourself in a subject and become a recognized world-leading expert. You’re pushing the boundaries of human knowledge, relatively free from everyday distractions and constraints. This is a remarkable endeavor, and if this doesn’t resonate with you, a PhD might not be the right fit.
The Disclaimer. It’s crucial to acknowledge the potential downsides and challenges. A PhD is a unique experience that requires a significant caveat. Expect to work intensely, especially before paper deadlines. You need to be prepared for periods of stress and possess the mental fortitude to handle pressure. There will be times when you lose track of the days, subsisting on leftovers from the lab kitchen. You might find yourself exhausted and alone in the lab on a sunny weekend, scrolling through social media posts of friends enjoying exotic vacations funded by their significantly larger salaries. You might have to discard months of work while striving to maintain your mental well-being. You may grapple with the realization that your research paper has few citations, while friends are launching exciting startups featured in tech publications or developing products used by millions. You might experience identity crises, questioning your life choices and the use of what might be considered your prime years.
Therefore, you should be confident in your ability to thrive in an unstructured environment driven by research and scientific discovery. If you’re uncertain, it’s wise to lean towards caution. Ideally, gain some research experience as an undergraduate through a summer program before committing to a PhD. In fact, a primary reason research experience is highly valued in PhD admissions is not just the research itself, but the fact that it helps students understand what they are getting into.
Let me be clear: this post isn’t intended to persuade anyone to pursue a PhD. I’ve simply outlined common considerations. The majority of this guide focuses on tips and strategies for navigating the PhD journey once you decide to embark on it, as we’ll explore below.
Finally, a thought I’ve heard is that a PhD is only worthwhile if you aim for academia. Considering the points above, I’d argue that a PhD has significant intrinsic value – it can be an end in itself, not just a means to an academic job.
Gaining Admission: References are Key. So, you’ve decided to apply. How do you get into a good PhD program? The primary factor is straightforward: strong letters of recommendation. The ideal letter comes from a well-known professor and reads something like: “X is in the top 5% of students I’ve ever worked with. They are proactive, generate their own ideas, and successfully execute them.” The least effective letter is along the lines of: “X took my class and performed well.” A research publication from a summer research program is a major advantage but not strictly necessary if you have compelling letters. Note that grades are less crucial, though you want to avoid exceptionally low grades. As an undergraduate, I mistakenly focused heavily on grades. That time would have been better spent on research or personal projects, ideally under the guidance of multiple mentors (as you’ll need 3+ recommendation letters!). Lastly, unsolicited and aggressive outreach to potential advisors is generally ineffective. They are often incredibly busy, and overly assertive attempts to impress them at conferences or via email can be off-putting.
Choosing a Program. Once admitted to several PhD programs, how do you make your choice? Joining Stanford is always an option (just kidding!). More seriously, your ideal program should: 1) Be a top-tier institution (not just for resume prestige, but because top schools attract other top minds, many of whom you’ll interact with and collaborate with), 2) Have several potential advisors whose research aligns with your interests. The “several” is crucial – it provides a safety net if things don’t work out with your initial advisor choice for various reasons beyond your control, such as a professor leaving or changing research direction, and 3) Be located in a desirable physical environment. New admits often underestimate this: you’ll spend 5+ years near the campus. This is a significant period, and your life will extend beyond research.
The Advisor Relationship: A Critical Dynamic
Image credit: PhD comics.
Image alt text: A humorous depiction from PhD Comics illustrating the complex and often symbiotic relationship between a PhD student and their academic advisor, highlighting the advisor’s guiding role in the student’s research journey.
Understanding the Advisor-Student Dynamic. Your advisor is exceptionally influential in shaping your PhD experience. It’s vital to understand the nature of this relationship. It’s a symbiosis: you have personal goals for your PhD, but advisors also have their own objectives, constraints, and career aspirations. Understanding your advisor’s incentives is crucial: tenure processes, evaluation criteria, funding sources, departmental politics, award acquisition, and the broader academic landscape, particularly how they gain recognition and respect within their field. This understanding can prevent many student-advisor friction points and facilitate effective planning. However, I don’t want to portray this relationship solely as a business transaction. Advisor-student relationships often evolve into lasting connections built on more than just career advancement.
Pre- vs. Post-Tenure Advisors. Advisors vary significantly. Tracking whether a potential advisor is pre-tenure or post-tenure is a useful heuristic (though exceptions exist). Junior faculty, pre-tenure, are often more hands-on and readily available as they are actively building their publication record to secure tenure. They may have stronger opinions on research direction, engage in detailed problem-solving, propose concrete ideas, and even review or contribute to code. This can be a more intense, hands-on experience, driven by the advisor’s need for publications and their incentive to push their students to work diligently. In contrast, senior, post-tenure faculty may lead larger labs and have broader commitments (committees, talks, travel), limiting their direct involvement in research details and student supervision. They tend to operate at a higher level, offering guidance like “explore this area further,” “talk to this person,” or “position your work this way.” In these cases, detailed guidance often comes from senior PhD students or postdocs in the lab.
Axes of Advisor Variation. Many other factors differentiate advisors. Some are informal, while others maintain a professional distance. Some heavily influence research details, while others are hands-off. Some focus on specific models and applications, while others prioritize research problems, regardless of the modeling approach. Managerially, some meet students weekly (or daily!), others are less frequent. Some respond to emails instantly, others may take days or weeks (or never!). Some impose strict work schedules (long hours, weekends), others are more flexible. Some provide ample resources and equipment, others expect students to make do with basic tools. Some generously fund conference travel, even without a paper presentation, others are more restrictive. Some are entrepreneurial and application-oriented, others favor theoretical research. Some encourage summer internships, others see them as distractions.
Finding the Right Advisor. How do you choose an advisor? The first step is in-person conversations. The advisor-student relationship is often likened to a marriage, requiring a good fit. Ensure you can communicate effectively and have a positive personal rapport. Crucially, assess where they fall on the “professor space” spectrum concerning the axes mentioned earlier, and whether there’s intellectual alignment in your research interests. Intellectual resonance is as important as management style.
Gathering Advisor References. Seek references on potential advisors. Talking to their current students is a valuable strategy. For candid feedback, informal settings are better than formal meetings. Students might hesitate to criticize their advisor directly in general inquiries, but they’re more likely to be truthful when asked specific questions like, “How often do you meet?”, or “How hands-on are they?”. Another tactic is to examine the career paths of their former students (often found on lab websites’ alumni sections), which statistically indicates potential outcomes for you.
Impressing a Potential Advisor. The advisor-student selection process is mutual. Advisors seek students they also want to work with. The ideal student is interested, passionate, self-directed, and proactive – someone who, after an initial suggestion, returns having not just met expectations but exceeded them, innovating in unexpected ways.
Consider the Entire Lab Group. Remember, you might meet your advisor weekly, but you’ll interact daily with their students, who will become your closest colleagues. You’ll likely collaborate with senior PhD students and postdocs, who act as mentors. Postdocs, essentially professors-in-training, are often eager to mentor to gain advising experience for their own academic job applications. Therefore, ensure the entire lab group is composed of people you respect, get along with, and can collaborate with effectively on research projects.
Research Topics: Navigating the Outer Loop
t-SNE visualization of a small subset of human knowledge (from paperscape). Each circle is an arxiv paper and size indicates the number of citations.
Image alt text: A visualization of academic papers from arXiv, illustrating the vast landscape of scientific knowledge and the interconnectedness of research topics, highlighting the depth and breadth of scholarly work.
Having joined a PhD program and found an advisor, the next question is: What research should you pursue?
The Meta-Problem Exercise. A PhD is both stimulating and frustrating because you’re constantly operating at a meta-problem level. You aren’t just solving predefined problems – that’s the inner loop. You spend most of your time in the outer loop, identifying worthwhile and solvable problems. You’re constantly envisioning potential projects, assessing their significance, impact, and feasibility. This can be mentally taxing, as you invest significant time without certainty about the problem’s validity or solution’s existence.
Developing Research Taste. Choosing research problems involves developing “taste,” a somewhat elusive academic concept. When you propose a problem to your advisor, their reaction – from disinterest and skepticism to excitement and engagement – reveals much. They’re rapidly assessing the problem’s importance, difficulty, novelty, historical context, and alignment with their ongoing research grants. Your advisor, likely a master of this outer loop, has a refined “taste” for research problems. During your PhD, you’ll develop this sense yourself.
Initially, my research taste was underdeveloped. Reviewing my early PhD notes, many problems I found exciting were, in retrospect, poorly conceived, intractable, or irrelevant. Practice and mentorship helped refine my judgment.
Here are some thoughts on developing research taste and identifying interesting problems:
A Fertile Research Area. Recognize that your PhD research will delve deep into a specific area, with papers building upon each other to form your thesis. Therefore, consider the long-term potential of your chosen area. Predicting the future is impossible, but you can often sense the potential for further research and exploration within a field.
Alignment with Advisor’s Interests and Strengths. Work within your advisor’s area of expertise. While some advisors allow tangential projects, you’ll benefit most by leveraging their knowledge and increasing their investment in your work. For example, consider your advisor’s “default talk” – their standard research presentation. If your work can add new, exciting, cutting-edge slides to this presentation, they’ll likely be more invested, helpful, and promote your research more actively. Their talks will also publicize your work.
Embrace Ambition: Sublinear Scaling of Hardness. People often mistakenly believe that a 10x more important problem is also 10x harder (or less likely) to solve. This is a fallacy. In my experience, a 10x more significant problem is at most 2-3x harder. In some cases, a 10x harder problem might even be easier to solve. Why? Because aiming for a 10x improvement forces you to think outside conventional boundaries, confront limitations, rethink fundamental principles, and innovate strategically. Aiming for a 10% improvement with hard work is achievable. But aiming for a 100% improvement, while still likely attainable, will require a fundamentally different approach.
Ambitious, Yet Actionable. Important problems abound, but not all make good PhD projects. Read Richard Hamming’s “You and Your Research,” which elaborates on this point:
If you do not work on an important problem, it’s unlikely you’ll do important work. It’s perfectly obvious. Great scientists have thought through, in a careful way, a number of important problems in their field, and they keep an eye on wondering how to attack them. Let me warn you, `important problem’ must be phrased carefully. The three outstanding problems in physics, in a certain sense, were never worked on while I was at Bell Labs. By important I mean guaranteed a Nobel Prize and any sum of money you want to mention. We didn’t work on (1) time travel, (2) teleportation, and (3) antigravity. They are not important problems because we do not have an attack. It’s not the consequence that makes a problem important, it is that you have a reasonable attack. That is what makes a problem important.
Becoming “The Person Who Did X.” Ultimately, a PhD aims to develop deep expertise and leave your mark on a field. The ideal outcome is to “own” a part of an important area, ideally one easily described. You want people to say, “She’s the person who did X.” If you can fill in that blank, you’ll have a successful PhD.
Valuable Skill Development. Recognize that a PhD makes you an expert in your chosen area. (As a fun aside, [5 years] x [260 workdays/year] x [8 hours/day] = 10,400 hours. If you believe Gladwell’s 10,000-hour rule, a PhD is precisely the time investment to become an expert). Imagine yourself five years hence, a world expert in this area (the 10,000 hours will ensure expertise regardless of academic impact). Will these skills be valuable and exciting for your future endeavors?
Problems to Avoid. Some problem types and papers are best avoided. “Incremental work” is a dreaded term in academia. Incremental papers enhance existing work, often by adding complexity, for marginal (e.g., 2%) improvement on benchmarks. Ironically, these papers have a decent chance of acceptance (reviewers struggle to find fatal flaws; they’re sometimes called “cockroach papers”). A string of these might create a sense of productivity, but they rarely become highly cited or impactful. Avoid projects based on “the next logical step that no one has done yet” or aiming for “an easy poster.”
Case Study: My Thesis. To illustrate, consider my PhD journey. My entire thesis was based on work done in the final 1.5 years. It took considerable time in the “meta-problem space” to find an exciting research direction. The preceding two years were spent exploring 3D (Kinect Fusion, 3D meshes, point cloud features) and video-related topics. Then, in my third year, I randomly visited Richard Socher’s office at 2 AM on a Saturday. Our conversation about interesting problems led me to realize the potential of his work on images and language (though this area’s history is much longer). I couldn’t foresee all subsequent papers, but it felt promising: a fertile area (many unsolved problems, grounding language in images), exciting and important, easily explained, at the edge of feasibility (Deep Learning was emerging), datasets were becoming available (Flickr8K had just been released), aligned with Fei-Fei’s (my advisor) interests, and even without major breakthroughs, I’d gain valuable experience optimizing deep networks, transferable elsewhere. A wave of “checkmarks” clicked into place. The next day, I proposed this area to Fei-Fei, who enthusiastically approved and guided me within the space (e.g., she encouraged image-to-sentence generation, while I was initially focused on ranking). I’m pleased with how things developed. In short, I spent two years navigating the outer loop, searching for a direction. Once heuristics aligned, I committed.
Resistance to Advisor Input. Remember, advisors are not infallible. I’ve witnessed and heard of instances where, in hindsight, advisors made incorrect recommendations. If you feel strongly about a different path, have the courage to sometimes diverge from advisor advice. Academia generally values independent thinking, though advisor responses vary. I know of cases where this divergence was highly successful and also personal instances where it was not. For example, in my first year, I disagreed with Andrew Ng’s advice, pursued a problem he was less enthusiastic about, and, unsurprisingly, he was right, and I wasted months. You win some, you lose some.
Beyond the “Game.” Challenge yourself to see a PhD as more than a sequence of papers. You’re not just a paper writer. You’re a member of a research community, aiming to advance the field. Papers are a common means, but look beyond the established academic “game.” Think independently, from first principles. Do things others should but don’t. Step off the prescribed treadmill. I tried this throughout my PhD. This blog is an example – it communicates ideas beyond papers. The ImageNet human reference experiments – I felt it crucial to establish baseline human accuracy on ILSVRC, so I spent weeks evaluating it. Academic search tools (arxiv-sanity) – frustrated by inefficient literature discovery, I created and maintain the site, hoping it helps others. Teaching CS231n twice – I invested more effort than rationally advisable for a PhD student focused on research, because I felt the field was hindered if people couldn’t efficiently learn and enter it. These endeavors likely came at the cost of standard academic metrics (h-index, top-venue publications), but I pursued them anyway, would do so again, and encourage others to do the same. To add nuance and acknowledge differing viewpoints, I know this perspective is contentious and many colleagues disagree.
Writing Papers: Mastering Academic Communication
Image alt text: The LaTeX logo, representing the typesetting system widely used in academia for preparing technical and scientific documents, especially research papers, emphasizing the importance of effective written communication in scholarly work.
Writing effective papers is a fundamental survival skill in academia, akin to fire-making for early humans. Crucially, understand that academic papers are a specific genre with conventions regarding format, flow, structure, language, and expected statistical rigor. Looking back at my early drafts is often painful – they were quite poor. There’s a significant learning curve.
Reviewing Papers: Learning from the Negative Examples. Trying to improve paper writing by only reading good papers is like learning binary classification solely from positive examples. Exposure to bad papers is equally valuable. Paper reviewing provides this. With typical conference acceptance rates around 25%, most papers you review are not strong. Analyzing weak papers reveals common pitfalls: unclear writing, undefined variables, vague introductions, premature dives into detail. This helps you avoid these mistakes in your own writing. Journal clubs are another valuable experience. Hearing experienced researchers critique papers provides insights into how your own work will be evaluated.
Getting the “Gestalt” Right. I recall being impressed by Fei-Fei’s (my advisor) paper reviewing efficiency. Presented with four papers I had reviewed over hours, she quickly flipped through each for about 10 seconds and accurately identified the one I accepted and the three I rejected. She relied on the gestalt of the papers, a powerful heuristic. As you become more experienced, your papers will develop a characteristic look. Introduction ~1 page. Related work ~1 page, dense with citations (but not overcrowded). A compelling “pull figure” (page 1 or 2) and system diagram (page 3), professionally designed, not in MS Paint. Technical sections with mathematical notation. Results tables with numbers, some bolded. An additional analysis experiment. Exactly 8 pages (the page limit), not a line less. Mastering this “gestalt” is crucial, as many researchers use it as a cognitive shortcut in evaluating your work.
Identifying the Core Contribution. Before writing, pinpoint the single core contribution of your paper. Emphasize “single.” A paper isn’t a random collection of experiments. It sells one novel, non-obvious idea. You must argue its importance, novelty, and then experimentally validate it in controlled settings. The entire paper is surgically focused on this core contribution, avoiding fluff or extraneous additions. In an early paper on video classification, I mistakenly included two contributions: 1) video convnet architectures and 2) an unrelated multi-resolution architecture for minor improvements. I reasoned that the latter might be interesting and that more contributions are better. This was incorrect. The minor contribution diluted the paper and was distracting. Similarly, my CVPR 2014 paper presented two models: ranking and generation. In retrospect, these should have been separate papers. Combining them was more historical accident than rational choice.
Paper Structure. Once you have your core contribution, a standard paper structure exists: Introduction, Related Work, Model, Experiments, Conclusions. When writing introductions, I find it helpful to outline a coherent narrative in LaTeX comments, then fill in the text. I structure paragraphs around a single point, stated in the first sentence and elaborated in the rest. This aids reader skimming. A good flow is: 1) X (define if needed) is an important problem. 2) Core challenges are A and B. 3) Prior work addressed these with Y, but has limitations Z. 4) We propose W (?). 5) W has properties P and Q, and experiments show results R and S. Adapt this structure, but these core points should be clear. The paper is surgically centered on your contribution. Challenges listed should be directly addressed by your work. Maintain structure throughout the paper. For example, in the model section: 1) Clearly state the section’s purpose. 2) Explain core challenges. 3) Describe baseline or prior approaches. 4) Motivate your approach. 5) Describe it in detail.
Breaking Structure Conventions. Feel free to add creativity. Razavian et al.’s 2014 paper (https://arxiv.org/abs/1403.6382) uses a student-professor dialogue introduction – clever and engaging. Papers from Alyosha Efros are often playful and excellent examples of engaging writing. See his paper with Antonio Torralba: “Unbiased look at dataset bias.” FAQ sections, perhaps in appendices, can also work well.
Avoiding the “Laundry List” Mistake. A common mistake is the “laundry list” structure: “Here’s the problem. To solve it, we do X, then Y, then Z, then W. Here are the results.” Avoid this. Justify, motivate, and explain each step. Why do X and Y? What are alternatives? What have others done? “This is common practice (cite)” is acceptable. Your paper isn’t a report, a chronological account of experiments, or a latex transcription of your notes. It’s a focused, processed discussion of a problem, your approach, and its context. It should teach colleagues something, justifying your steps, not just describing them.
Language Choices. Develop a vocabulary of “good” and “bad” words for academic writing, particularly in fields like machine learning and computer vision. Avoid passive, weak verbs like “study” or “investigate.” Use active, stronger verbs like “develop” or “propose.” Don’t present a “system” or, worse, a “pipeline”; develop a “model.” Don’t learn “features,” learn “representations.” Avoid incremental, weak terms like “combine,” “modify,” or “expand.” These can signal incremental work and potential rejection.
Internal Deadlines Two Weeks Prior. While not universal, Fei-Fei’s lab enforces an internal deadline two weeks before external submission. A 5-page draft with final experiments (even preliminary results) undergoes internal review mirroring the external process. This practice is incredibly valuable. Forcing a full paper draft almost always reveals crucial experiments needed for paper coherence and argument strength.
Jennifer Widom’s “Tips for Writing Technical Papers” is another excellent resource.
Writing Code: Best Practices for Research Software
Image alt text: A close-up of computer code on a screen, symbolizing the essential role of software development in research, particularly in computational fields, and highlighting the need for well-written, documented, and shareable code.
Much of your time will involve implementing your ideas, often through extensive coding. While coding isn’t unique to academia, here are some key points:
Release Your Code. Surprisingly, you can publish papers without releasing code. Incentives to avoid release are strong: it’s work (research code is often messy, requiring cleanup), it can be intimidating to expose your coding skills, maintaining code and answering user questions is ongoing, and bugs might be revealed, potentially undermining results. However, these are precisely the reasons to release your code. Fear of public scrutiny motivates better coding habits (saving time in the long run!), forces improved engineering practices, encourages thoroughness (e.g., unit tests), increases research impact (more citations), and serves as a useful, verifiable record of your work. When releasing code, use Docker containers to minimize dependency-related user issues.
Think of Your Future Self. Document your code meticulously, for yourself. Returning to your codebase months later (e.g., for camera-ready revisions), you’ll likely feel completely lost. Develop a habit of creating detailed readme.txt files in all repositories, documenting code functionality, usage, etc., as notes for your future self.
Giving Talks: Engaging Your Audience
Image alt text: A speaker presenting at an academic conference, emphasizing the importance of effective oral communication skills for researchers to disseminate their work, engage with the community, and enhance their professional impact.
You’ve published a paper, and it’s an oral presentation! You have a few minutes to present to a large audience. What makes a great talk?
Talk Goals. A common misconception: the talk’s goal is to summarize your paper. Incorrect. That’s a secondary objective. The primary goals are: 1) Excite the audience about the problem (they must appreciate it to care about your solution!). 2) Teach the audience something (give them a taste of your insight/solution; don’t shy from discussing related work). 3) Entertain (or they’ll check social media). Ideally, attendees leave thinking: “Wow, I’m in the wrong research area,” “I must read that paper,” and “This person deeply understands this field.”
Talk “Do’s.” Several elements improve talks: Pictures! People love visuals. Use videos and animations sparingly, as they can distract. Make talks actionable – suggest what the audience can do after your talk. Give live demos if possible; they’re memorable. Develop a broader intellectual context for your work. Craft a narrative – people love stories. Cite extensively! It takes little slide space to credit colleagues, pleases them, and reflects well on you, showing humility and awareness of prior and parallel work. Even cite related work presented at the same conference and briefly advertise it. Practice! First alone, then with labmates/friends. This almost always reveals flaws in narrative and flow.
Talk “Don’ts.” Avoid text-heavy slides. Minimal or no bullet points. Speakers sometimes use bullet points as prompts, but slides are for the audience, not the speaker. Put speaker notes in the notes section. Avoid complex diagrams. Your “simple” diagram is likely not simple or easily interpretable to a first-time viewer. Audience bit bandwidth is limited.
Cautionary Notes: Result Tables. Avoid dense result tables showing your method is better. You got a paper accepted; your results were likely decent. These are often boring unless the numbers reveal something insightful (beyond “our method is better”) or if there’s a dramatically large, noteworthy improvement. If including results/graphs, build them gradually with transitions; don’t present everything at once and linger on a single slide for minutes.
The Bored/Confused Pitfall. Designing talks that effectively teach is challenging. A common failure is talks that are boring initially (too general overview) and then confusing (too technical second half), leaving the audience with no takeaways. Identify if your talk risks this pattern.
Time Management Pitfall. Many speakers overspend time on introductory parts (often less engaging) and then rush through the most interesting results, analysis, or demos. Avoid this.
Formulaic Talk Pitfall. While personal preference, I favor non-formulaic talks that challenge conventions. Outline slides are often tedious. It’s like saying, “This movie is about a ring of power. Chapter 1: hobbit gets ring. Chapter 2: journey to Mordor. Chapter 3: ring destroyed. Chapter 1 begins.” No! Use outline slides sparingly, mainly for very long talks (30+ minutes) to re-anchor audience attention if they drift, but judiciously.
Observe and Learn. The best way to improve talks (like paper writing) is to consciously analyze great (and weak) speakers. Don’t just passively listen; analyze, deconstruct, and learn. Pay close attention to audience reactions. Notice when people disengage (e.g., during complex tables, many check phones). Develop an internal classifier of audience-disengaging events and avoid them in your talks.
Attending Conferences: Beyond the Presentations
Image alt text: A bustling poster session at an academic conference, showcasing the vibrant exchange of ideas and research findings in a less formal setting, emphasizing the social and networking aspects of academic gatherings.
Regarding conferences:
Attend. Conference attendance, especially top conferences in your field, is crucial. If funding is limited and your advisor can’t cover expenses (e.g., if you’re not presenting), consider self-funding (typically around $2000 for travel, accommodation, registration, and food). It’s an investment in becoming part of the academic community, meeting peers, and discussing research trends. Despite the lone-wolf scientist stereotype, research is highly social. You build upon others’ work, collaborate, and write papers for this community. Furthermore, fields have “tacit knowledge” not always in papers – emerging research areas, interesting papers, insider perspectives, historical context, methods that work in practice, etc. Becoming part of this community and accessing this “hive mind” is invaluable and enjoyable – first to learn, and then to contribute.
Talk Selection: Speaker-Centric. When choosing talks, prioritize speakers over topics. Some individuals are consistently excellent speakers (a learned skill). Attending their talks, even on tangential topics, is often rewarding.
Hallways are Where the Action Is. Innovation, especially in Machine Learning, moves faster than conference cycles. Many presented papers are already “old news.” Conferences are primarily social events. View the hallway as a primary, unscheduled event. Poster sessions are also valuable for discovering interesting, potentially overlooked papers and ideas.
The three stages of a PhD: 1) You look at a paper’s references and haven’t read most. 2) You recognize all the papers. 3) You’ve shared a beer with the first authors of all the papers.
Closing Thoughts: Integrity and Long-Term Success
I recall Sam Altman (YC) saying there are no shortcuts or cheats in building a startup. Long-term success isn’t achieved by gaming the system or faking appearances. This applies to academia. Focus on good research and advancing the field. Gaming proxy metrics is unsustainable. Academia is surprisingly small and interconnected. Shady tactics to inflate your academic record (excessive self-citation, redundant publications, unchanged paper resubmissions, omitting baselines, etc.) will eventually be detrimental.
Ultimately, it’s simple: Do good work, communicate it effectively, and recognition and positive outcomes will follow. Enjoy the journey!
EDIT: HN discussion link.
