For the last 6 years I’ve dabbled in many things, but only one thing consistently. I’ve been running this weekly event, DLCT, rarely missing any week, for 6 years straight. The format of DLCT is simply, “talks”: a speaker comes to talk about a deep learning paper (usually one of their own), and engage with the audience (of a size between 40 to 80 on average) for an hour.
Amongst the many ecosystems our work—whether it is scientific papers or tech products—can inhabit, the “events ecosystem” is one of more challenging ones. At least, it has been for me. Events often feel secondary, existing primarily to promote another product or idea, operating in a landscape that’s fiercely competitive. Any time the Internet latches onto a new buzzword, a swarm of events suddenly appear, each one of them emblazoned with that very term. The ultimate currency in this ecosystem is “attention.” And that drives a relentless pursuit of big names, deep-pocketed sponsors, and flashy marketing tactics.
As a typical cynic, I often deliberately distance myself, and the things I do, from the prevailing trends—what everyone else seems to be doing. For one, DLCT sessions are usually not recorded, so I’m not even trying to maximize the reach. For another, I made DLCT radically open to people who don’t usually get invited to speak, through an anonymous nomination form on the website. This way, I minimize my personal bias, and give everyone a chance.
“Giving everyone a chance” is almost my motto these days.
Still, a few stars have come to the show, some before they were stars (examples: Chis Olah in 2019, Dan Hendrycks in 2020, Jonathan Ho in 2021, Lilian Weng in 2021, Rohan Anil in 2021, Jeff Clune and Jim Fan in 2022). A few papers presented there only later became wide adopted standards (InstructGPT in 2020, NeRF in 2021, ConvNext in 2022).
Fun aside: I just did a rough count, We’ve had DLCT sessions 239 times.
Another fun aside: I’ve myself presented (or jointly presented) a total of 11 times at DLCT, but never once was it about my own paper. I have been really careful about minimizing personal bias.
Through managing the DLCT stage, I’ve seen a wide spectrum of speakers. Some are naturally gifted communicators, while others are still honing their craft. It’s a natural outcome for the radically removed barrier of entry: the quality of talks does vary. But I’d say not below any common academic standard. It’s hard to do technical talks well: they are just…well, too technical. The academic community doesn’t always incentivize strong presentation skills. I mean, the ability to give compelling talks is always a plus, but it hasn’t yet reached the top of the priority list for most researchers.
Around Year 3, I began keeping a note of those who are good speakers. This is purely for my own amusement, a personal memory logging rather than a professional endeavor, and I have no intention of ever sharing it publicly. I don’t ever aspire to become a speaking-skills certifier of some sort. Even though the machine learning mindset often dictates that improving a skill requires first finding a way to evaluate and measure that skill, which then allows you to “hillclimb” it, it would be a dystopian nightmare if we ever decide to put out a leaderboard for people, and rank them on skills such as “speaking.”
All that the note says is: at some specific snapshot in some people’s lives, some were really good at this specific thing. It’s a time bound label, which means after passing that labeling timestamp, it shouldn’t mean anything. I see everyone, me included, to be on a path of trying to get better; where we eventually end on that path is merely a function of the time remaining on our hands, and the amount of energy we put behind it. I feel privileged to have the chance of witnessing snapshots, to be exact, 239 of them, in the growth paths of all the speakers that have come to the show, and have no doubt that they will only grow better after this.
There are also different kinds of goodness (and my simple note does make a distinction of that): there’s the highly polished type, from usually someone that has been speaking all their life, and there’s the what I call “endearing” type, which I adore it so much that I made a special label for it in my note.
The endearing type is one that while their content, language and delivery may not be perfect, you can clearly see their heart behind it — that they really take this chance of communication seriously, in that they put the audience first, before themselves. Like any good science, a good talk is where you take yourself and your ego out of it.
And this is the sole reason I feel like writing this essay. After six years of curating a weekly talk series, and bearing witness to nearly 240 presentations (if we take into account all my conference attendances and meetings during my day jobs, this number would be thousands), I’ve come to see clearly that in scientific speaking, our heart determines the end point much more than our muscles. The most captivating talks aren’t defined by eloquence or stagecraft, but by a genuine will to connect with the audience. Empathy, not oratory, is the true cornerstone of effective communication.
While polished language and confident delivery—which can be well practiced and improved over time—can certainly enhance a presentation, they are ultimately secondary to the speaker’s willingness and ability to connect, to understand and empathize with their listeners’ needs and interests. I call the former “muscle skills”—skills like eloquence and fluency that you can measure, benchmark, and hillclimb, similar to how you enhance your muscles by lifting weights, and the latter “heart skills”—skills that equal to how much heart and soul you put behind an effort, which no one can measure but yourself. Technically it’s not really a skill but an inclination, since it’s hard to acquire from imitating or practicing, but comes from one’s innate desire, the fuel to one’s fire.
This realization is rather consoling to someone like me. Because the muscle skills are more correlated with privilege (whether you happened to be raised and trained in an environment that emphasizes those skills), than the heart ones. And I would love to live in a world where the guiding principle is: “If you want to get better at something, you will get better. Because the very ‘want’ is all that matters.”
I don’t know if that principle is true in general, or in other fields, but I’m at least seeing the truth of it in this very narrow walk of life that is giving machine learning paper talks.
Let me clarify what I mean by empathy in the context of scientific talk giving. It means that your one and only goal is for the audience to understand your content. Not your charisma, not your stage presence, not your self-image. Just the clarity and impact of your message. It’s about stripping away any ego-driven motivations—how you look, sound, or how much applause you garner.
More concretely, the “endearing” type I noted usually do these few remarkable things:
- Going slow, especially in the begining when a concept is first introduced
- Take long pauses to allow proper digestion of a new concept, and to allow questions
- Not the least caring about finishing their slides, but rather the extent of people understanding what’s already presented
- Not glossing over, well, anything
I think we are entering an era where talks will lose their relevance if they are just an information dump, rather than a chance of connection. And I eagerly welcome this world where the ones with the bigger hearts, not the ones with slicker tongues, will rise to the forefront.