Episode 4 [Unedited]

Podcast: Play in new window | Download (Duration: 26:31 — 36.4MB) | Embed

Subscribe: Apple Podcasts | Google Podcasts | Stitcher | Email | TuneIn | RSS

This is episode is the full, unedited interview with Mandy Korpusik. If you haven’t listened to the fully produced episode yet, we strongly encourage you to do so before listening to this one. They’re shorter in length and much more refined.

EPISODE CREDITS

Guest Starring Mandy Korpusik, Phd Candidate at MIT CSAIL
Produced & Hosted by Adam Greenfield
Executive Produced by Patrick Yurick, Instructional Designer – MIT OGE
Executive Produced by Heather Konar, Communication Director – MIT OGE

Special thanks to the following editors who provided us invaluable feedback that aided in the development of this show:

Christopher O’Keeffe, Co-Founder of Podcation

Kristy Bennet, Manager – MIT Women’s League

Jennifer Cherone, Phd Candidate – MIT Burge Laboratory

Erik Tillman, Phd, Formerly of the Kim Lab & Currently A Fellow at Vida Ventures, LLC

The Great Communicators Podcast is a part of Gradcommx. Gradcommx, targeted at enhancing research communication, is the first offering of Gradx – a professional development project created for the graduate student population at the Massachusetts Institute of Technology by the Office For Graduate Education.

MUSIC & SOUNDS

“Divider” by Chris Zabriskie is licensed under Attribution 4.0 International License (http://freemusicarchive.org)

EPISODE SCRIPT

• Print The Script Here

ADAM GREENFIELD

Hello, Adam Greenfield here, host of The Great Communicators podcast series. What you’re about to hear is the full, unedited interview with one of the guests we spoke with. If you haven’t listened to the fully produced episode yet, I definitely encourage you to do so before listening to this one. They’re shorter in length and much more refined. You can find them all at gradx.mit.edu/podcasts.

The idea behind these longer, unedited conversation is to give you an opportunity to hear the entire talk, warts and all. This is not only a fun way to hear the full flow of the conversation but it also emphasizes the importance of the points made in the shorter, produced episodes, which again, can be found at gradx.mit.edu/podcasts.

Thanks for listening and enjoy the conversation.

Patrick Yurick: Alright, so, interview question one is what is your name and what is your field of study?

Mandy Korpusik: My name is Mandy Korpusik and I’m studying natural language processing in the computer science and artificial intelligence laboratory.

P: What is that, specifically, what do you do?

M: So, a lot of my lab mates work on speech recognition but I’m interested in what’s coming back from the speech recognizer. So if someone is using Siri, kind of understanding what they’re asking and how to respond back to them. And specifically I’m working on a diet tracking application. I’m trying to understand if someone describes their meals naturally, what the foods are that they ate and then what the nutrition facts are.

P: How’d you get involved with that?

M: Actually, we’re collaborating with some nutritionists at Tufts University. So they had this idea, oh, if we can use speech and language technology, maybe it would be easier for people, like their patients, to track food that they’re eating.

P: So is this an algorithm you’re working on so that would work with other different speech recognition software?

M: Actually, I’m not really dealing with the speech recognition stuff so it’s basically the technology that understands a spoken or written meal description. Like, I had a bowl of oatmeal and an apple and a glass of milk. And then it figures out that the foods are the oatmeal and the apple and the milk and maps that to the USDA database to get the nutrition facts.

P: What’s the most complicated part of all of that?

M: Right now, the most interesting part I’ve done is this deep learning algorithm using a neural network to map the meal description to the matches in the database. But in my future work, I think being able to respond to someone and give personalized nutrition advice or actually interact with them through dialogue is going to be really tricky. And quantities. Figuring out exactly how much someone ate of something is kind of a challenge, too.

P: Why?

M: People don’t really know how much they ate and when they describe things it doesn’t really map directly to the database quantities. Like, if someone said, I had a bowl, they might not know how many cups that is. Or if they say, oh, I had some butter on my toast. It’s like, how many tablespoons is that, exactly?

P: So it sounds like there needs to be smart bowls and utensils and-

M: Something to do the quantities. Or maybe if you take a photo and use computer vision to estimate quantity from the photo.

P: Well, yeah, it kinda makes sense. So, Weight Watchers for years has been making tons of money off this portion control stuff and the way they do it is they put- they sell portioned things to the- this is how they make money. You can either go to the store and buy the Weight Watchers proportioned foods or they sell the cups that you can measure and a scale.

M: Interesting. Hmm.

P: Basically what they do is provide you the information as like, oh, if you eat X amount of food it counts as this many points. So they’re working on a points conversion system. My mom did it for while so I know how they did it.

M: Oh, ok.

P: It’s the reverse of what you’re trying to solve, which is what I prefer. Like, I would rather have an app that’s like, hey man, that’s a little too much cereal this morning. I’d rather it be like knowing what I’m doing as opposed to me having to remember to count my calories. Because when I’m hungry the last thing I’m thinking about is how many calories you’re putting into an app. So it just made me think that smart utensils that connect to your phone could keep track of that. Oh, we noticed you ate a lot of red meat this week. But that doesn’t sound like it would be a ton of work, given some of the sensors but I don’t know how smart utensils would work. I know a bowl could easily count weight but it couldn’t distinguish, oh, you’re eating Cheerios. Maybe it could just ask you.

>>>>>

P: Let’s talk about communication. What kind of struggles do you have when it comes to communicating about your science to- I mean, you’re speaking next week.

M: So I guess something I have to keep in mind is how technical the audience is because I did a presentation earlier this year at CSAIL, it was at the CSAIL Research party, and it was a five minute talk and you know your audience is in computer science but not necessarily in your field. So they were kind of helping me get rid of all the super technical stuff. I had this really detailed neural network model and they were like, that’s way too much. And all these numbers on the table and they’re like, no, you don’t want that in this talk. It should be like, here’s Alice and Bob and they’re trying to track their dietary intake and this is how they were doing it before, they were typing it into Google, and typing into the USDA food database browser and there’s a long list of options and it’s painful for them. So really highlighting the motivation. And then for my talk I prepared my slides and I sent it to my advisor and he’s like, um, this isn’t the CSAIL Research party, you should get rid of Alice and Bob and make it more technical. Because I was like, oh, it’s supposed to be 17, 18 minutes but I have too many slides so I hid the ones that are the super technical stuff on how I perform this ranking in the top matches in the database. And he said we should definitely unhide those slides, bring it back in. So then I took one of the figures that was kind of simplified and then I replaced it with one of the ones that was in the paper and more technical.

P: It sounds like you went through kind of slingshot feedback. One piece of feedback was to make it more simple and one feedback was to make it more complicated. Have you gotten better over time at figuring out which to do and which situation- or do you still rely on feedback from people?

M: I guess now hopefully I’ll be able to kind of modify it better given the audience but yeah, I’m still learning. And even outside of MIT, just talking to people, like my friends or relatives, I’ve noticed it has to be completely different than how I talk to my lab mates, for example. My relatives say, oh, I didn’t understand at all what you said you worked on. Can you explain it again? And I kind of go back and rethink, ok, what did I say to them and how can I make it easier for people who aren’t technical to understand. So I think I’ve gotten better at that now.

P: When you come up with an idea for something, do you have to kind of filter it through, how do I communicate this to somebody at some point? Is that part of it? Because it sounds like you’re in this hotbed where so much cool stuff is happening with the research around you and different people are working on different things. What’s your first reaction when you come up with an idea for how to solve a problem? Do you figure how to communicate it or do you try to do it yourself and then try to figure out how to talk to somebody about it?

M: I guess, yeah, when I’m thinking about an idea, like a solution to some problem, I’m thinking about how to approach it from a technical standpoint, and then later, once I’ve kind of worked through it, then I would frame it depending on the audience.

P: How much do you get tripped up even talking to your lab partners and different people you’re collaborating with? Do they all really understand everything that you know?

M: Not necessarily because they might be doing speech recognition and I might be doing natural language processing and there’s a difference even in between us. But I think they understand pretty well.

P: Have you figured out shortcuts to get them up to speed so they can understand how you’re coming at something because you’re doing speech recognition vs image processing?

M: I think most of the models and solutions we use are pretty similar between the two even though it’s pretty different tasks. So if I just stick to talking about the implementation instead of- I don’t know. It’s not super hard with my lab mates. They can pretty much understand as long as I explain it clearly enough.

P: Yeah. I’m interested because one of the things we found when talking to a lot of the- I’ve just been thinking about communication for a year on this project- and jargon and technical speak tends to also be shorthand. When it’s effective it’s awesome because it can mean you can jump somebody up to speed much faster because you don’t have to explain, I don’t know, quantum entanglement. Like, if you know what that means and I say that and I know what that means, we can move on to the next idea quicker versus if I say quantum entanglement and you’re like, huh? I’m like, alright, I have to explain that, or I have to think to myself, is it worth it to explain that or is there only a small part of that I need to be able to move on to the next thing.

M: Ok, that’s probably what it is. We understand all of these key words and these different models and so we can just use the jargon with each other.

P: What are some of the key words that are common in the work that you’re doing?

M: So, I might say I’m using a CNN architecture, which means a convolutional neural network and so they understand the acronym and they know what this network is and they know how it’s different from other possible neural network architectures. Or LSTM, long short term memory neural network. And so they know what I mean when I say the CNN performed better on my task then the LSTM. Or if I say I’m using a word embedding layer as the bottom layer of my neural network then they know what the word embedding is.

P: LSTM. How would you explain that if, say, I was your grandmother?

M: Ok. I guess I would say it’s some machine learning model that can take some sort of input and predict some sort of answer as the output given- uh, this is very difficult. It’s something remembers what it has seen in the past and it kind of remembers that information as it’s moving forward through time and it relies on that as context.

P: So it’s a machine that remembers what it’s done?

M: Yeah.

P: Interesting. Ok. I was only asking because I think it’s interesting to hear the difference. And also the point, it is a struggle to think about- and I’ve encountered this a lot with other grad students- you haven’t thought of it in that simple term but I also think it’s kind of difficult because if I dumb it down too much, it’s not actually accurate, right? But if I’m just trying to construct an understanding to get somebody from point A to point B, or Z, I’m going to skim over the middle of the alphabet because it’s not really important to remember all the technical stuff for them to just understand why it’s important. Which is interesting. What’s you’re initial reaction after listening to the podcast?

M: I enjoyed the podcast. I liked the interview with Professor Chomsky because for one thing I’m studying natural language processing so I took linguistics classes for my minor. So Chomsky is really famous in our field so that was cool. And even the stuff about tying it to education because I’m interested in education. I taught for the Women’s Technology Program last summer and so one of the things they kind of trained us on was how he said you’re leading your students to discover knowledge instead of covering the knowledge. I liked that. It kind of resonated with what I already studied. And the story from the other professor, when they were giving a talk and there were 5000 people and then 2000 got up and left in the middle I thought that was kind of powerful so it’s good to remember to motivate your work for your audience.

P: Did anything you hear change your concept of audience or add a new layer of understanding to the concept of audience that you were thinking before you listened to it?

M: I’m not sure.

P: Ok. Was there anything that specifically stood out to you that was helpful?

M: Yes. I liked the part where it was towards the end where it was saying if you notice your audience is kind of falling asleep or not paying attention anymore, then you can just try to hook them again and say, you know what’s really interesting? Ok, I should try that because when I’m teaching, that’s when it happened the most often, when the high school girls were just kind of getting tired of my 9:30 lecture and falling asleep. Yeah, so. I liked that.

P: >>>>(Patrick’s Randy Pausch story) So when I was first becoming a teacher, especially with high school students, I studied charismatic speech, public speakers, because high school students will literally be like, I don’t know what you’re talking about but this seems really exciting. I know a lot about high school students but adults? Not so sure. Especially the MIT crowd. I did my first speech here to scientists and it was scary.

M: Why?

P: Because all the beats that I use with teachers- so teachers, their values are driven by emotions. So for instance, when you’re talking to teachers you need to talk about students.  The reason why is because teachers will only value what you have to say if you can demonstrate you care about the same thing that they care about, which is the students. If you start talking to teachers about how they need to be teaching first, they won’t listen to you because they’re like, you’re just another dude coming in to tell me how me how to do my job like people do all the time. So with teachers you have to demonstrate that you care about the students before you start talking about the craft of teaching. Scientists, you have to talk about the data, because they don’t care about anything else. You cannot have any conclusions that you have arrived to that are declarative. That’s a big thing. All declarative statements have to have room for- you have to demonstrate that you’re open to disagreement about any conclusions that you may have found about your work. The curiosity factor is really big but none of those declarative statements can come out unless it’s backed up by data. They don’t care about emotions at all.>>>>>>>> Are there any techniques, tricks, advice, etc. that you have used in the past when it comes to connecting with your audience?

M: I think actually one thing Tony had taught me in my class was using hand gestures to kind of emphasize your point. So when I’m explaining why it’s painful to use my fitness pal and you have to scroll through a long list of database options and I kind of use my hand to show it and my voice to emphasize things. And when I’m talking to my uncle, I’ve learned people don’t really like- he said, some people in your generation speak too quickly. So I try to slow down and enunciate for my uncle.

P: But do you think that that would- basically, earlier, you said that would change based on who you’re talking to, right?

M: Yeah, I guess so. These things are probably good for anyone. But they come into my mind with specific audiences.

P: And Tony wasn’t in this podcast so can you just mention who he is and just that you took a class in communication with Tony?

M: So I took this public speaking class my first semester here with Tony Eng and he was teaching us how to communicate.

P: Was that your first communication course?

M: I think so.

P: What struck you from that experience? Has anything stuck with you? Was it weird to go through the class?

M: It was interesting. I think the main thing that stuck with me was just being comfortable speaking in front of other people. That was really the main thing. We would get up and then we would have to say stuff off the top of our head, like the words that come into our mind or just stare into each other’s eyes. And these exercises, I think, were super helpful, just becoming comfortable being in front of a large audience.

P: Do you have any advice or anything to other grad students, based on any experiences or stories you have or either you’ve done something right or you wish you’d done something different to engage an audience?

M: I think I’ve learned over time to really stick with pictures in my slides and avoid text. And I think that’s my main thing when I’m trying to prepare a presentation. The more pictures I have, the easier it is for them to connect what I’m saying with the visual and really kind of understand what I’m trying to get across. And reading text does not go over well. People seem bored when I do that.

P: For next week, are you trying anything new?

M: I guess I haven’t really- no, I haven’t planned anything new. Although, I might see if I can do something to perk people up if they’re falling asleep or they look bored.

P: Are you nervous about anything for next week?

M: I feel pretty good but I’m mostly nervous because it’s my first talk at a big conference. I’ve given one technical talk before but that was a small workshop so this is a bigger one. Oh, answering questions. That’s what also makes me nervous. Yeah, because I’m like, what are they going to ask me? Are they going to ask something really tough?

P:>>>> I wonder, too, in your field, have you observed anybody who has done that really well, that answered random questions that you can model after?

M: Even just watching my advisor’s really helpful because he’s so good at giving talks and he’s also good at giving questions. The professors.

P: What are some things he does or your professors do when answering questions? Do they just seem like they know everything?

M: Maybe it’s just they’re really confident and they know their topic really well but he’s probably just developed that with experience.

LISTEN TO MORE EPISODES