<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Amir-Hossein Karimi on Medium]]></title>
        <description><![CDATA[Stories by Amir-Hossein Karimi on Medium]]></description>
        <link>https://medium.com/@amirhkarimi?source=rss-38efa6030da3------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*eVe6nrikgy20-7ST.</url>
            <title>Stories by Amir-Hossein Karimi on Medium</title>
            <link>https://medium.com/@amirhkarimi?source=rss-38efa6030da3------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 20:42:33 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@amirhkarimi/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Learning to Think for Yourself in the Age of AI]]></title>
            <link>https://amirhkarimi.medium.com/learning-to-think-for-yourself-in-the-age-of-ai-f783962c1c8b?source=rss-38efa6030da3------2</link>
            <guid isPermaLink="false">https://medium.com/p/f783962c1c8b</guid>
            <category><![CDATA[phd]]></category>
            <category><![CDATA[independent-thinking]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Amir-Hossein Karimi]]></dc:creator>
            <pubDate>Fri, 13 Nov 2026 19:35:01 GMT</pubDate>
            <atom:updated>2026-11-13T19:35:01.365Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*hMECVtdrgyChlW7lkaZByw.png" /></figure><p>AI can help you read a paper, write code, or come up with ten experiments by lunchtime. You still have to decide which question is worth asking, which experiment could answer it, and what to believe when the results come back.</p><p>Those decisions come up whenever you’re trying to figure something out for yourself. Research makes them especially visible because you have to explain why you made them and show what happened when you put them to the test.</p><h3>Five skills to develop</h3><ol><li><strong>Identify a problem worth solving.</strong> Read the literature, understand what people have already tried, and ask what we’d learn if the problem were solved.</li><li><strong>Formulate it.</strong> Turn a broad interest into a question precise enough to investigate. Make your assumptions explicit.</li><li><strong>Break it down.</strong> Figure out which pieces you need to solve and in what order. A good first step often tests the assumption the rest of the project depends on.</li><li><strong>Solve and evaluate.</strong> Choose a method, test it, and work out what the evidence supports. Decide what to do when the results surprise you.</li><li><strong>Communicate the contribution.</strong> Explain what you found, why it matters, and where your conclusions stop.</li></ol><p>You can practice these skills on a small project. Take a result you couldn’t explain, read enough to turn it into a question, and design a test that could distinguish between two plausible explanations. Write up what you learned and ask someone which part of your reasoning they’d challenge. You can do this at work, with collaborators, or on a project of your own.</p><h3>What a PhD offers</h3><p>A PhD gives you sustained time to practice these skills, with an advisor and colleagues who can challenge your reasoning. Early on, I might help define a project and meet frequently. As you take ownership of more decisions, we usually meet less often. When I consider whether a student is ready to go before a PhD committee, I ask whether they can independently demonstrate all five skills and defend the choices they made along the way.</p><p>A good thesis solves a concrete problem and opens up several others. Once you’ve answered one question carefully, you can often see consequential questions that were hard to formulate when you started. The world has plenty of problems worth investigating. Part of becoming a scientist is learning to recognize the ones where you can make a contribution.</p><p>I don’t think a PhD is for everyone. It’s a commitment of several years, and it matters what you’d otherwise do with that time. If you’re considering one primarily to reach a particular job, ask whether that job requires a doctorate and what other routes could get you there. If what you want is to become an independent scientist, the training itself can be worth those years. A PhD is one setting for that work.</p><h3>How to prepare for graduate research</h3><p>Use your statement of purpose to practice the first steps. Imagine you’ve just joined a lab and you’re bringing an idea to your first research meeting. What question would you propose? What makes it worth asking in light of what’s already known? What would you try first, and what would that tell you?</p><p>My rule of thumb is to spend about two-thirds of the statement making that case and one-third explaining how your past work shaped the question and approach. I go into more detail in the companion post, <em>How to Write a Good Statement of Purpose</em>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f783962c1c8b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Researchfluencers]]></title>
            <link>https://amirhkarimi.medium.com/researchfluencers-7b1a5491e2f2?source=rss-38efa6030da3------2</link>
            <guid isPermaLink="false">https://medium.com/p/7b1a5491e2f2</guid>
            <category><![CDATA[social-media]]></category>
            <category><![CDATA[academia]]></category>
            <category><![CDATA[research]]></category>
            <category><![CDATA[influencers]]></category>
            <category><![CDATA[twitter]]></category>
            <dc:creator><![CDATA[Amir-Hossein Karimi]]></dc:creator>
            <pubDate>Wed, 02 Jun 2021 11:25:56 GMT</pubDate>
            <atom:updated>2021-06-02T11:28:35.190Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1000/1*cAfkQJEkdEeTg61uidlPMQ.jpeg" /><figcaption>Source: <a href="https://www.strategy-business.com/blog/Make-the-most-of-your-influencers?gko=b2f55">https://www.strategy-business.com/blog/Make-the-most-of-your-influencers</a></figcaption></figure><p>What am I describing?</p><p><strong>Step 1.</strong> Select a catchy title, maybe something that’s controversial.</p><p><strong>Step 2.</strong> Write a captivating first few lines, enough so to hook the reader to continue.</p><p><strong>Step 3.</strong> Add a carefully crafted image, with eye-watering color blends and amplified contrast that’s sure to capture attention.</p><p><strong>Step 4</strong>. Once the audience is hooked, tell a relatable backstory, exude confidence, and show that you know your stuff.</p><p><strong>Step 5.</strong> Now, sell your brilliant ideas, designs, and dreams.</p><p><strong>Step 6.</strong> Finally, conclude with an exciting vision of the future, and maybe plug some of your other brilliant insights.</p><p>Did I just describe writing an Instagram post, a Twitter thread, or an academic paper?</p><p>Exactly.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/582/1*jOGPwQZYbMim0ZOuTZutjw.png" /></figure><p>Before there were follower counts and number of re-tweets, academics had citations and h-indices.</p><p>Before there was “going live” and paid “audio-only spaces”, academics had invited talks and summer schools.</p><p>I posit that academics and researchers are perhaps the first influencers. If I may, “researchfluencers”.</p><p>Just as researchfluencers compete for funding to deliver novel insights and push the boundaries of understanding, ever-connected YouTube/Twitter/ Snapchat/Instagram/ClubHouse/TikTok/LinkedIn/Facebook/Substack/ Medium/Reddit (WSB anyone?) influencers compete for your attention by A/B testing the “Algorithm” [1] to push their content to a broader audience.</p><p>The difference lies not in the action or motive, but in the audience and medium.</p><p>Whereas the former writes about the design of a complex system through a peer-reviewed publication, the latter tweets about a deep philosophical realization via 280 characters, or posts an Instagram “Reel” of their creamed coffee as they were philosophizing (yes, it’s a word).</p><p>Just as we value the impact of highly influential academic papers [2,3], one cannot deny the impact of a well-placed meme on social change [4,5].</p><p>At the end of the day, we are all social creatures who find validation and opportunity through sharing and exchanging. New mediums simply allow more of us to easily express ourselves to broader audiences and with fewer barriers of entry (e.g., higher education, gate-kept opportunity, etc.).</p><p>This is not to say that traditional researchfluencers don’t use modern-day influencing tools to interact with and broaden their audience. Over the past year and accelerated by the pandemic, many instructors have partially or fully migrated their teaching material to platforms such as YouTube, academic venues such as large machine learning conferences use SlidesLive to host recordings of papers and keynote presentations, and Twitter has become a go-to source of not only staying up to date with the latest work but also disseminating ones own [7,8,9].</p><p>As a final thought, I leave you with this tweet [6] which has sturred a lot of reflection in me. The gist of it is that Mark Rober, an ex-NASA engineer who literally helped put the Curiosity Rover on Mars, has since started a YouTube channel with the slogan of “new ideas, science, and engineering”. He puts out YouTube videos each month for his follower base of 18 M+ people, which garner millions (if not 10&#39;s of millions) of views within the first week of release. The same Mark Rober co-authored a peer-reviewed publication proposing tubing configuration and design for the Mars Science Laboratory Descent Stage, which had a whopping 323 views since 2009. Both statistics are out-of-this-world ;)</p><p>[1] <a href="https://sproutsocial.com/insights/social-media-algorithms/">https://sproutsocial.com/insights/social-media-algorithms/</a></p><p>[2] <a href="https://www.altmetric.com/top100/2020/">https://www.altmetric.com/top100/2020/</a></p><p>[3] <a href="https://www.natureindex.com/news-blog/google-scholar-reveals-most-influential-papers-research-citations-twenty-twenty">https://www.natureindex.com/news-blog/google-scholar-reveals-most-influential-papers-research-citations-twenty-twenty</a></p><p>[4] <a href="https://time.com/5922998/the-most-impactful-memes-of-2020/">https://time.com/5922998/the-most-impactful-memes-of-2020/</a></p><p>[5] <a href="https://mountainscholar.org/bitstream/handle/10217/183936/Huntington_colostate_0053A_14303.pdf">https://mountainscholar.org/bitstream/handle/10217/183936/Huntington_colostate_0053A_14303.pdf</a></p><p>[6] <a href="https://twitter.com/joshmnicholson/status/1266523445938065409?s=20">https://twitter.com/joshmnicholson/status/1266523445938065409?s=20</a></p><p>[7] <a href="https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0229446">https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0229446</a></p><p>[8] <a href="https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3278109/">https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3278109/</a></p><p>[9] <a href="http://blogs.lse.ac.uk/impactofsocialsciences/2017/12/04/academic-journals-with-a-presence-on-twitter-are-more-widely-disseminated-and-receive-a-higher-number-of-citations/">http://blogs.lse.ac.uk/impactofsocialsciences/2017/12/04/academic-journals-with-a-presence-on-twitter-are-more-widely-disseminated-and-receive-a-higher-number-of-citations/</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7b1a5491e2f2" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Interviewing for a PhD in ML? Here are some sample coding questions.]]></title>
            <link>https://medium.com/data-science/interviewing-for-a-phd-in-ml-here-are-some-sample-coding-questions-a0d228155f3f?source=rss-38efa6030da3------2</link>
            <guid isPermaLink="false">https://medium.com/p/a0d228155f3f</guid>
            <category><![CDATA[interview]]></category>
            <category><![CDATA[software]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[coding]]></category>
            <dc:creator><![CDATA[Amir-Hossein Karimi]]></dc:creator>
            <pubDate>Wed, 24 Mar 2021 14:30:00 GMT</pubDate>
            <atom:updated>2021-03-31T10:01:48.325Z</atom:updated>
            <content:encoded><![CDATA[<h4>A guide for interviewers and interviewees alike.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-gZHs5d4Yv8Mi6a1yU1jDg.png" /><figcaption>Source: <a href="https://pixabay.com/vectors/job-interview-career-conference-156130/">https://pixabay.com/vectors/job-interview-career-conference-156130/</a></figcaption></figure><p>Machine learning PhD interviewees are typically assessed on various dimensions: interpersonal, presentation quality, technical foundations, and coding ability.</p><p>During my PhD, I’ve had the opportunity to conduct coding interviews for several candidates for the PhD programs at <a href="https://is.mpg.de/">MPI-IS</a> and <a href="https://learning-systems.org/">CLS</a>. Prior to this, I first started hosting coding interviews at Facebook Inc. and learned quite a bit from my mentors there (thanks, KK!). In this blog post, I present a sample set of questions that I use to guide the candidate through the process.</p><p>The questions progress from simple and immediately solvable to more nuanced. The questions are also designed in such a way so as to have both straightforward solutions (which may not be correct), as well as more involved (but not necessarily efficient) solutions, leaving ample room for creative discussion.</p><p>Candidates are encouraged to write pseudo-code or use the syntax of the language they’re more comfortable with; naturally, given that this is a machine learning interview, most candidates directly write in python.</p><p>Besides writing code, the interview covers run-time and memory analysis, touches on commonly used data structures (hash maps and binary trees), and basic combinatorics.</p><p>Needless to say, most candidates have been performed quite well on this set of questions, have mostly finished in time (~40 minutes), and seem to have enjoyed the format (I hope to receive feedback from those whom I’ve interviewed!)</p><p>Let’s get started.</p><h3>Question 1</h3><pre>def is_palindrome(s):<br>    ‘’’return True if `s` and its reverse are the same, False otherwise’’’</pre><p>This question is rather straightforward, and without exception, all candidates solve it in the first 5 minutes. The purpose has two purposes. Firstly, it quickly gauges a candidate so that one can adjust the next questions accordingly if this is difficult. More importantly, this question is a quick and easy win for the candidate, boosting their confidence and relaxing the conversation for them.</p><p>A couple of potential solutions may look as follows:</p><pre>def is_palindrome(s):<br>    for i in range(len(s//2)): <br>        if s[i] != s[-i-1]: <br>            return False <br>    return True</pre><p>Or</p><pre>def is_palindrome(s):<br>    return s == s[::-1]</pre><p>And finally, I ask about the run-time, which is O(n) for both solutions above. Again, quick and easy wins help the candidate to think comfortably, as they would in a non-interview setting. Sometimes, when the candidate uses built-in methods, or libraries (e.g., to revert the string), this hints at experience, which shows itself further in the next questions.</p><h3>Question 2</h3><pre>def is_anagram(a, b):<br>    ‘’’return True if string `a` and `b` are permutations of one another, False otherwise’’’</pre><p>This is one of my favorite questions, as it has multiple solutions with varying levels of implementation and run-time complexity. This encourages creative thinking with the candidate and often leads to an interesting discussion.</p><pre>def is_anagram(a, b): <br>    for c in a: <br>        if c not in b: <br>            return False <br>    return True</pre><p>This is a simple and elegant first solution and works well for some pairs of inputs, e.g., {123, 312}. But it also wrongfully returns True for the {123, 3124} pair (try it out!).</p><p>So how do we proceed? A seemingly simple fix would be to assert equal lengths for a and b, resulting in:</p><pre>def is_anagram(a, b): <br>    assert len(a) == len(b) <br>    for c in a: <br>        if c not in b: <br>            return False <br>    return True</pre><p>But upon closer inspection of this solution (a top candidate is expected to challenge their own solution and test corner cases), we can see that this solution would not suffice as it fails for other pairs, e.g., wrongfully returning True for {122, 123}.</p><p>Then the candidate notes (or is reminded) that the order of input arguments, a and b, is important. A simple next step would be to repeat the nested for loop in the other direction as well, resulting in:</p><pre>def is_anagram(a, b): <br>    assert len(a) == len(b) <br>    for c in a: <br>        if c not in b: <br>            return False <br>    for c in b: <br>        if c not in a: <br>            return False <br>    return True</pre><p>At this point, I ask the candidate about run-time. Given the nested for-loop structure, this is O(n²) run-time. I also ask whether they think the solution is complete, to which most respond with a counterexample to the above. In particular, the next potential pitfall is how to deal with repetitions. For instance, even with the assertions and bi-directional nested-for-loop, the above code fails for {1233, 1123}.</p><p>There are, once again, multiple ways to resolve this. Once a candidate realizes that they should think about a solution to handle repetitions, most start ideating about ways to amend the above code with individual character counters; I quickly guide them away from implementing this from scratch as it would be time-consuming and not so fun! Instead, I encourage them to think about data structures that could store such counters automatically.</p><p>Here most candidates point to a hash map (dictionary) that would keep track of all unique characters in addition to their repetition count. The solution then relies on constructing two dictionaries, namely, char_counts_a and char_counts_b for each input string respectively, and then comparing the key-value pairs of these two dictionaries. This solution has a trade-off requiring a bit more memory and the previous solution, however, it has a run-time of O(n) and successfully handles strings with non-unique characters.</p><p>At this point, while we have identified a solution to the question, and while the answer seems to be the most efficient possible (at least to the best of my knowledge), the interview can continue by gauging other aspects of the candidate’s knowledge.</p><p>Here I asked the candidate to think creatively about alternative ways they would solve this question. The following are two bonus approaches that we discuss time permitted in the interview.</p><p>Almost always, the candidate smiles when asked to think creatively, perhaps a sign that they’re having fun :) strongly encouraged!</p><p>First, I asked the candidate to revisit the definition of an anagram pair. I remind the candidate that by definition anagram means that the two strings are permutations of one another. What do you think a solution based on permutations will look like? Hint: run-time: O(n * n!). The big-O run-time is the same, with and without repetitions, another question worth discussing to gauge basic combinatorics.</p><p>Second, I ask the candidate to consider alternative approaches that they may have learned (e.g., sort) in an introductory CS course on algorithms. What do you think a solution based on permutations will look like? Hint: run-time: O(n * log n).</p><p>It goes without saying that each interview proceeds uniquely, and the answering order above is not representative of how different candidates think and approach a solution. The difficulty of the interviewer role (yes, interviewers have it hard too!) is to have many potential solutions simultaneously in mind so as to help the student navigate to a solution. Assessing candidates is not always straightforward either, as it is not uncommon for candidates to perform poorly on straightforward questions, but quickly see nuances and corner cases in more involved questions. Rarely can one conclusively assess coding ability with one question, and thus I find it doubly important to ask many small questions.</p><h3>Question 3</h3><pre>def fibonacci(n):<br>    ‘’’return the nth number (0-index) in the Fibonacci sequence (1,1,2,3,5,8,13), e.g., fibonacci(4) would return 5.’’’</pre><p>A common (and good) first solution suggested by candidates involves a for-loop over the Fibonacci sequence digits, and looks something this:</p><pre>def fibonacci(n): <br>    fib_nums = [1,1] <br>    for i in range(2,n): <br>    fib_nums.append( <br>        fib_nums[-1] + <br>        fib_nums[-2]<br>    ) <br>    return fib_nums[-1]</pre><p>This solution has a run-time of O(n). A not-so-uncommon alternative involves recursion, as follows:</p><pre>def fibonacci(n): <br>    if i == 0 or i == 1: <br>        return 1 <br>    return fibonacci(n-1) + fibonacci(n-2)</pre><p>While this solution looks elegant and clean, it, unfortunately, suffers from an exponential run-time. Working through this run-time exercise with the candidate provides a lot of good signals. Strong candidates are not only quick at identifying the call stack deficiency, they are also comfortable working with data structures (here, binary trees) to derive what ultimately amounts to a O(2^n)run-time. The most experienced candidates also suggest ways to improve, commonly resorting to storing already computed values of the Fibonacci sequence, either explicitly as in the first answer, or implicitly through caching (e.g., using the <a href="https://www.python-course.eu/python3_memoization.php">@Memoize</a> decorator in python).</p><p>Importantly, as is the case in many interviews, arriving at the final answer is only one part of the evaluation. Even if a candidate does not immediately arrive at solution, it is an opportunity for the interviewer to gauge the candidate’s receptiveness to new ideas and a general curiosity for learning new concepts. In my experience, candidates almost always walk away pleasantly engaged with this question.</p><p>Finally, there are existing <a href="https://en.wikipedia.org/wiki/Fibonacci_number#Closed-form_expression">closed-form expressions</a> for the Fibonacci numbers which rely on the golden ratio and some involved arithmetic. This is a final (cherry-on-top, but not-so-necessary) signal, whereby candidates can compare real-world tradeoffs between the first O(n) solution which can compute the answer using 8-bit memory, and the closed-form O(1) which should run using floating-point operations.</p><h3>Conclusion</h3><p>I hope you learned something from these sample interview questions. Of course, these questions are not restricted to a machine learning PhD interview, and can be used more broadly for other technical coding-based interviews. For ML positions specifically, I would argue that seeing beyond initial solutions and being proficient at gauging and contrasting solutions on various dimensions is an essential skill, as we seldom come across easy all-encompassing solutions. These questions (try to) assess this ability, along with testing rudimentary knowledge of rapid prototyping skills any PhD candidate should have. Most importantly, the questions are constructed in a way to guide the interview in a calm and relaxing setting for the candidate, starting from simple confidence-building to more difficult creative-thinking-requiring discussion.</p><p>Feel free to use these questions, modifications thereof, or the general framework in your next interview. If you have feedback, do let me know <a href="https://twitter.com/heikalkhan">@heikalkhan</a>!</p><p>Now that I’ve shared my secret questions, I have to go look for new questions. Adios!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a0d228155f3f" width="1" height="1" alt=""><hr><p><a href="https://medium.com/data-science/interviewing-for-a-phd-in-ml-here-are-some-sample-coding-questions-a0d228155f3f">Interviewing for a PhD in ML? Here are some sample coding questions.</a> was originally published in <a href="https://medium.com/data-science">TDS Archive</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[PhD Advice]]></title>
            <link>https://amirhkarimi.medium.com/phd-advice-333d8cfb1d4b?source=rss-38efa6030da3------2</link>
            <guid isPermaLink="false">https://medium.com/p/333d8cfb1d4b</guid>
            <category><![CDATA[academia]]></category>
            <category><![CDATA[academic-life]]></category>
            <category><![CDATA[phd]]></category>
            <category><![CDATA[phd-life]]></category>
            <dc:creator><![CDATA[Amir-Hossein Karimi]]></dc:creator>
            <pubDate>Sun, 16 Aug 2020 20:05:06 GMT</pubDate>
            <atom:updated>2020-08-30T10:40:21.603Z</atom:updated>
            <content:encoded><![CDATA[<p>Recently a junior PhD student asked my advice on best-practices, attitudes, and general outlook when pursuing their PhD. Their impression was that I, as somebody with more PhD-years under his belt, would have valuable insight to share. While this is debatable, they wrote to me a few hours after our conversation, thanking me profusely for the comments. Below I share some of what we discussed in hopes that you may find this useful as well. If not, feel free to unfollow me on Twitter 🐦 <a href="https://twitter.com/heikalkhan">@heikalkhan</a>. ;)</p><p>Before I start, I want to emphasize that each person experiences a PhD differently and often under very different circumstances. This could be the result of personal learning style or communication habits: some are more visual learners and may prefer working on collaborative group projects, others may enjoy spending countless hours exploring the mathematical foundations of a problem alone, whereas others foster their most creative ideas during mid-day coffee chats on unrelated topics ☕️. Experiences may vary as a result of relationship factors such as the 1–1 interaction with their supervisor, and how well they get along with other lab members, the department, and the university. Of course, a PhD is (should be?) only one of a number of engagements that the candidates takes part in, and it’s silly to assume that personal and professional life will always be aligned; the unpredictability of factors that may be out of our control is what makes this chapter even more interesting.</p><p>With this disclaimer out of the way, here are some tips that worked for me personally. To better understand my background and motives, briefly: I am a PhD candidate in computer science, with a background in software engineering, and a drive to have (hopefully positive) impact through robust AI 🧠, scalable software 💻, and entrepreneurship💡.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/909/1*jVoP3megwQ0H3XFwgUeADQ.jpeg" /><figcaption>source: <a href="https://www.tharawat-magazine.com/family-business-ownership/psychological-ownership-next-generation/">https://www.tharawat-magazine.com/family-business-ownership/psychological-ownership-next-generation/</a></figcaption></figure><h3><strong>Own your PhD</strong></h3><p>Decide early on, ideally before starting, why you want to do a PhD, and what you hope to gain from it. There are many great articles that discuss this detail [<a href="https://gradadmissions.mit.edu/blog/why-would-you-want-do-phd">1</a>,<a href="https://www.quora.com/What-are-three-good-reasons-for-doing-a-PhD">2</a>,<a href="https://medium.economist.com/why-doing-a-phd-is-often-a-waste-of-time-349206f9addb">3</a>,<a href="https://phdadvice.carrd.co/">4</a>], here are some ideas to tickle your brain:</p><ul><li>Do you hope to broaden your knowledge in a specific field (e.g., astrophysics)?</li><li>Is there a specific or personal problem (e.g., fairness in machine learning) that you would like to tackle?</li><li>Are you seeking a specific post-graduation career (e.g., in academia or at a research lab) a PhD is a (necessary) requirement?</li><li>Do you want to live in a particular city or country (e.g., to live in France and travel Europe), and continuing your education in this region is a way to accomplish that?</li></ul><p>Having a personal answer to this question is very important. You should seek an answer without self-judgment and by honestly reflecting on your intentions. Inevitably, there will be times during your PhD when communication, experiments, results, or publications do not go as planned, or they may outright fail. There may be situations when you must make difficult choices about opportunities, perhaps as a result of conflict between personal and professional goals (e.g., should I pursue a 6-month fellowship while I am pregnant?). You will meet many people, with a diverse flavor of impressive backgrounds, and imposter syndrome always lurks in the shadows [<a href="https://www.nature.com/articles/d41586-019-03036-y">5</a>,<a href="https://tophat.com/blog/imposter-syndrome-academia/">6</a>,<a href="https://portlandpress.com/biochemist/article/42/3/62/225249/Coping-with-imposter-syndrome-in-academia-and">7</a>]. It is precisely at these critical points in time where awareness of your goals and intentions will help you think clearly about the next steps, what matters to you personally, and why.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*DjmnncXnA_aZlRl_nfzcpw.jpeg" /><figcaption>source: <a href="https://theclaimsdesk.co.uk/support-network/">https://theclaimsdesk.co.uk/support-network/</a></figcaption></figure><h3><strong>Build your support networks</strong></h3><p>Whatever your motivation, make sure to communicate them with your supervisor or academic mentor early on. While they may disagree with your motives, or suggest detours, or not be able to help you directly, as somebody more academically experienced than you, it’s important to hear their feedback. At worst, if you don’t reach a consensus due to irreconcilable differences in objectives or style, you can seek another supervisor. At best, you agree on the goals and in turn, set a research plan that works along those lines. In any case, a sensible supervisor would try to assist you through this. Mentoring you through your PhD towards milestones really is part of their job description [<a href="https://exp.med.ubc.ca/faculty-2/responsibilities-for-research-supervisors/">8</a>,<a href="https://www.imperial.ac.uk/media/imperial-college/administration-and-support-services/registry/qa/public/good-practice/roles-and-responsibilities/Research-Degree-Supervisor.pdf">9</a>]. On this note, please remember that both you and your supervisor are intelligent agents that understandably and on an average act in your best self-interest. Thus, it’s quite likely that you won’t be able to always get the support you need from your mentor alone.</p><p>Beyond your supervisor, it is important to realize that you are not alone in this. Perhaps the people that share and understand most of your going through are your lab members. Making an effort to be part of group lunches, after-work outings, or late-night coffee runs before a deadline are all great ways to learn more about your colleagues. Not only is this an opportunity to learn about their research and expand on your hard technical skills, but also through these interactions you will come to learn about similar struggles and further develop soft communication and empathy skills. I have personally found these interactions pretty rewarding, and preventative in terms of falling into my own lonely bubble concerning my research.</p><p>Finally, and rather counterintuitively, professional ambitions often require that you take some time away from them. Striking the right “professional/personal” balance is an individualistic question, and also rather difficult to answer. At critical points in time, this balance may be grossly violated (e.g., before submission deadlines, or due to unexpected personal hardship). You may find that this balance is thrown off at other times also (maybe you’ve fallen out of your running habits due to consecutively bad weather, and never got back?). Other circumstances might blur the lines entirely (a pandemic?). Regardless of the balance that works best for you, it may be helpful to think of a PhD as one of a number of engagements in your life, and hence why I prefer not referring to this as “work/life” balance. Building a support network outside of your professional circle, with whom you foster your personal interests, may not only reduce the risk of burn-out [<a href="https://www.nature.com/articles/nj7654-375a">10</a>], it also nurtures creativity in your professional career, and it would allow for a more diverse development of your interests, ultimately resulting in more joy. As my father always says, “everything in moderation’’ (not too much, not too little).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1000/1*fl3UE7eeKumalOxNOT_GGg.png" /><figcaption>source: <a href="https://hellodr.com.my/product/kardiamobile/">https://hellodr.com.my/product/kardiamobile/</a></figcaption></figure><h3><strong>Hello, Dr. X</strong></h3><p>Whereas the previous two points apply generally to most profiles, this last point is more tailored towards individuals who seek to continue a post-PhD academic career. Admittedly, the message may be controversial or even narrow-minded; it’s the best I know of today and is a work-in-progress. Feedback is warmly appreciated!</p><p>If you plan on continuing in academia, or any field for that matter, then it’s important to be aware of some common metrics used to evaluate candidates. For better or for worse, publication records, academic talks, short course offerings, and general service to the community are commonly used yardsticks to assess candidates at various levels of their career: from becoming a PhD candidate, to defending, to writing grants, and applying for tenure.</p><p>In this game, proposal writing is a common requirement also. Here candidates are typically asked to write about 1–3 sub-fields towards which they have meaningfully contributed or plan to contribute. Essentially, the committee expects to see strong candidates to have not only a grasp, but a mastery of the literature, tools and methodology, and prospective ideas of said subfields.</p><p>Extrapolating backward, it seems reasonable to expect a senior near-defense PhD student to be somewhat of an expert (not necessarily world-class but not far away either) in at least the one area of research (likely the core idea of their thesis). One way to think of this is, when you graduate, you become Dr. X. E.g., Dr. microbiome, Dr. algorithmic recourse, Dr. machine learning fairness, or Dr. charm quarks (yes, that’s really a thing [<a href="https://en.wikipedia.org/wiki/Charm_quark">11</a>]).</p><p>What will you become? What is your X?</p><h3><strong>Summary</strong></h3><p>In summary, my personal experience has taught me that a PhD is a game of tradeoffs, and balancing them in such a way that works for you is what makes it a unique experience for each candidate. In your PhD, you will be challenged in ways you may not be ready for, and you will grow as a result. Having clear goals, building a support network (professional and personal), and focusing intently on the central topic of your thesis are three suggestions that may make navigating these waters a bit easier. With that, good luck, and I would be happy to hear your comments below.</p><h3>Acknowledgments:</h3><p>Special thanks to Miriam Rateike and Patrick Putzky for inspiring me to write this piece and challenging some of my more crazy ideas. :)</p><p>[1] <a href="https://gradadmissions.mit.edu/blog/why-would-you-want-do-phd">https://gradadmissions.mit.edu/blog/why-would-you-want-do-phd</a></p><p>[2] <a href="https://www.quora.com/What-are-three-good-reasons-for-doing-a-PhD">https://www.quora.com/What-are-three-good-reasons-for-doing-a-PhD</a></p><p>[3] <a href="https://medium.economist.com/why-doing-a-phd-is-often-a-waste-of-time-349206f9addb">https://medium.economist.com/why-doing-a-phd-is-often-a-waste-of-time-349206f9addb</a></p><p>[4] <a href="https://phdadvice.carrd.co/">https://phdadvice.carrd.co/</a></p><p>[5] <a href="https://www.nature.com/articles/d41586-019-03036-y">https://www.nature.com/articles/d41586-019-03036-y</a></p><p>[6] <a href="https://tophat.com/blog/imposter-syndrome-academia/">https://tophat.com/blog/imposter-syndrome-academia/</a></p><p>[7] <a href="https://portlandpress.com/biochemist/article/42/3/62/225249/Coping-with-imposter-syndrome-in-academia-and">https://portlandpress.com/biochemist/article/42/3/62/225249/Coping-with-imposter-syndrome-in-academia-and</a></p><p>[8] <a href="https://exp.med.ubc.ca/faculty-2/responsibilities-for-research-supervisors/">https://exp.med.ubc.ca/faculty-2/responsibilities-for-research-supervisors/</a></p><p>[9] <a href="https://www.imperial.ac.uk/media/imperial-college/administration-and-support-services/registry/qa/public/good-practice/roles-and-responsibilities/Research-Degree-Supervisor.pdf">https://www.imperial.ac.uk/media/imperial-college/administration-and-support-services/registry/qa/public/good-practice/roles-and-responsibilities/Research-Degree-Supervisor.pdf</a></p><p>[10] <a href="https://www.nature.com/articles/nj7654-375a">https://www.nature.com/articles/nj7654-375a</a></p><p>[11] <a href="https://en.wikipedia.org/wiki/Charm_quark">https://en.wikipedia.org/wiki/Charm_quark</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=333d8cfb1d4b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Machine Learning Security, and its interplay with Model Interpretability]]></title>
            <link>https://medium.com/data-science/machine-learning-security-and-its-interplay-with-model-interpretability-a13fa8362af8?source=rss-38efa6030da3------2</link>
            <guid isPermaLink="false">https://medium.com/p/a13fa8362af8</guid>
            <dc:creator><![CDATA[Amir-Hossein Karimi]]></dc:creator>
            <pubDate>Thu, 20 Jun 2019 21:28:50 GMT</pubDate>
            <atom:updated>2020-09-22T17:15:48.266Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*FAhR0rkfEHk6E5MaQKD19g.jpeg" /></figure><p>The development and deployment of machine learning present opportunities and challenges on many fronts: from the curation of high quality and relevant datasets, to the development of generalizable models, and finally to the deployment of said models. In the final camp, issues such as fairness, accountability, and transparency have been central to many recent research questions (see FATML community [1]), as well as interpretability and data privacy.</p><p>Model security is another important consideration. In practice, model security aims to prevent the theft of model parameters of a deployed model (accessible via an API) by an adversary, a loss which could have competitive/legal/robustness consequences. Such concerns affect both startups and companies, specifically the bigger players such as Clarifai, Google, Amazon, and others whose monetization strategy relies on amortizing model training cost over the lifetime of their API [2].</p><p>This tension is really fascinating when a company must decide between allowing access to a very powerful API (with detailed information on model outputs including class labels, confidence scores, etc) while risking the replication of their model by an adversary. Even when the compromising solution suggests that the API only provide limited information, regulations demand that a deployed model provide interpretable explanations (a legal requirement [3]), which may facilitate model stealing yet again [4].</p><p>Recently, my collaborators and I introduced MACE: Model-Agnostic Counterfactual Explanations for Consequential Decisions [5]. The motivation was to provide an all-encapsulating method to satisfy the legal requirement of interpretable explanations on various ML models and distance metrics (representing a diverse set of real-world applications). Next, I’m excited to explore the potential of counterfactual explanations for simultaneously providing high-quality explanations with provable guarantees, without disclosing too much about the internals of the model. Achieving this is necessary for the adoption of ML models by consumers, and by the law.</p><p>[1] <a href="https://www.fatml.org/">https://www.fatml.org/</a></p><p>[2] <a href="https://arxiv.org/pdf/1609.02943.pdf">https://arxiv.org/pdf/1609.02943.pdf</a></p><p>[3] <a href="https://arxiv.org/pdf/1711.00399.pdf">https://arxiv.org/pdf/1711.00399.pdf</a></p><p>[4] <a href="https://arxiv.org/pdf/1807.05185.pdf">https://arxiv.org/pdf/1807.05185.pdf</a></p><p>[5] <a href="https://arxiv.org/pdf/1905.11190.pdf">https://arxiv.org/pdf/1905.11190.pdf</a></p><p><strong><em>Note from Towards Data Science’s editors:</em></strong><em> While we allow independent authors to publish articles in accordance with our </em><a href="https://towardsdatascience.com/questions-96667b06af5"><em>rules and guidelines</em></a><em>, we do not endorse each author’s contribution. You should not rely on an author’s works without seeking professional advice. See our </em><a href="https://towardsdatascience.com/readers-terms-b5d780a700a4"><em>Reader Terms</em></a><em> for details.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a13fa8362af8" width="1" height="1" alt=""><hr><p><a href="https://medium.com/data-science/machine-learning-security-and-its-interplay-with-model-interpretability-a13fa8362af8">Machine Learning Security, and its interplay with Model Interpretability</a> was originally published in <a href="https://medium.com/data-science">TDS Archive</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>