Introduction To New Co-Host, Theresa Eimer

[00:00:00] Welcome to the AutoML Podcast. It's been a minute.

[00:00:04] So, I had many exciting things for us in the works

[00:00:08] before the Middle East decided to erupt in its latest configuration

[00:00:13] and that configuration fully absorbed me into it.

[00:00:17] And this wasn't the first time it's been happening

[00:00:20] but I promise to spare all of you the political details.

[00:00:23] I know this is the wrong podcast to airing this out

[00:00:26] but I should say that, like everybody else,

[00:00:29] I hope to see a day where all of this is behind us.

[00:00:33] As you might remember from one of the first episodes,

[00:00:36] one of the reasons I became drawn to AI and to AutoML in particular

[00:00:41] is because of the promise that I've been seeing in it

[00:00:44] and the proliferation of this technology

[00:00:47] and the attendant good that it can do.

[00:00:50] But I also saw the potential harm in it

[00:00:52] and I believed at the time and I still believe now

[00:00:55] that the public deserves greater exposure to what we're up to.

[00:00:59] I think some of this had already changed with ChadGPT.

[00:01:02] The public is becoming much more aware of these things

[00:01:05] and that we ourselves could also do well to speak openly

[00:01:08] about the work that we're doing

[00:01:10] and also to philosophize about its moral and ethical consequences.

[00:01:14] And so since the war started,

[00:01:16] I've really been trying to tackle those kinds of consequences more fully

[00:01:20] and I've been working on ways to try and leverage AI

[00:01:23] to facilitate better understanding between people

[00:01:26] who live in different political realities,

[00:01:29] in particular the different ideological realities

[00:01:31] that we're seeing in the Middle East.

[00:01:33] This is likely a quixotic project,

[00:01:36] which again won't be my first,

[00:01:39] but I did decide to try and pursue this line of activism

[00:01:42] which I've been toying with on and off for many years now,

[00:01:46] but I wanted to try and pursue it this time much more fully.

[00:01:49] So this explains the relative quiet

[00:01:51] you've been hearing on this front for the last few months.

[00:01:54] Sorry about that.

[00:01:56] But quiet you will hear no more

[00:01:58] and this is thanks to an inspiring new co-host of the show,

[00:02:01] Teresa Aima, which I know that some of you already know.

[00:02:05] Among other things, Teresa is a PhD student in Marius's lab

[00:02:09] at the Institute of AI at Leibniz University

[00:02:12] and she'll be taking the lead on hosting many of the future episodes of the show

[00:02:16] and I'm thrilled to have her.

[00:02:19] She's already recorded a bunch of conversations,

[00:02:21] which are going to air shortly.

[00:02:23] So you definitely want to stay tuned for those.

[00:02:26] And in this conversation today, Teresa and I will chat about some of the work

[00:02:30] she's been doing, she'll introduce herself.

[00:02:33] And Teresa, it is an honor and a privilege to be taking the stage with you.

[00:02:38] Let's get started.

[00:02:41] You have a tattoo, right?

[00:02:44] Yeah, I got one on my forearm.

[00:02:47] I got this when I was relatively freshly 18 or 19 or something.

[00:02:53] Just a cool design and a line of poetry I liked

[00:02:56] and then I accumulated a few, I have some flowers on my back.

[00:02:59] I actually got a souvenir tattoo from ICML last year.

[00:03:03] It's some flowers.

[00:03:05] I saw it while walking to the conference each day.

[00:03:08] I was walking by a tattoo studio and I saw this motif all the while

[00:03:12] and then after the conference was over, I stayed for a few more days.

[00:03:16] I just walked in there and was like, can I have those flowers on my arm?

[00:03:19] Any lines of code? Do you have any?

[00:03:21] Or an equation? Not yet.

[00:03:24] No, not yet. Let's see if that's still happening.

[00:03:28] I thought the other day, so I don't have any tattoos,

[00:03:31] but I thought, I think I saw somebody with a spider tattoo

[00:03:35] on their, it's like a spider on their wrist.

[00:03:38] And I thought to myself, oh no, it was a spider on their neck

[00:03:41] and I don't like spiders at all.

[00:03:44] But I was thinking, could it be that like you get totally desensitized

[00:03:48] to it if you just see it all the time?

[00:03:50] So then I thought, oh man, maybe I should just get like

[00:03:53] an incredibly gnarly spider on my wrist.

[00:03:56] You know, and just see it all the time.

[00:03:59] My wife will kill me Adam, my wife will kill me.

[00:04:03] You're living in Germany now, right?

[00:04:05] Yes, I am.

[00:04:06] Did you grow up there? How did you get there?

[00:04:09] I mean, I grew up in Bavaria. Some people claim that's Germany.

[00:04:13] What was your journey into the auto ML world?

[00:04:17] That was a bit more complicated, I guess.

[00:04:20] I originally started studied biochemistry

[00:04:24] and just noticed that, you know, lab environment isn't my thing.

[00:04:27] So I switched to computer science, but the theoretical computer

[00:04:30] science for all my bachelor's, most of my master's.

[00:04:34] And then I didn't even take an auto ML lecture.

[00:04:38] I took a reinforcement learning lecture.

[00:04:40] And the only people doing reinforcement learning

[00:04:42] were the auto ML group.

[00:04:44] So then I started doing RL with them and that just ended up being

[00:04:48] 50% auto ML and yeah, all downhill from there one could say.

[00:04:53] And you decided then to, well, did you want to stay in academia

[00:04:57] at least for your PhD? What did you have in mind?

[00:05:01] Well, I didn't even apply to PhD positions.

[00:05:04] I actually had gotten an industry job or well,

[00:05:08] I had an interview and they told me I could have the job

[00:05:11] and then I defended my master's thesis the day after,

[00:05:14] which is when Mario, so my PhD supervisor asked me if I wanted

[00:05:18] to do a PhD with him since he was moving to become a professor

[00:05:22] at that exact time. And I just thought, yeah, sounds nice.

[00:05:25] Get to see a different place, get to work on this topic.

[00:05:29] I've been enjoying why not?

[00:05:31] So there's never really been a plan about academia

[00:05:33] or industry and there still really isn't.

[00:05:35] Yeah. And so what kind of research do you do now?

[00:05:38] Yeah, it's smack in the middle between auto ML and reinforcement

[00:05:43] learning. So we're really trying to make reinforcement

[00:05:46] learning work out of the box, which it really doesn't.

[00:05:50] So if you've ever tried to apply reinforcement learning

[00:05:52] to a new problem that's involved because it can be brittle.

[00:05:56] It can be really hard to get it to work.

[00:05:59] And yeah, that's something that auto ML is good at fixing

[00:06:02] the principle, but reinforcement learning is also a bit more

[00:06:05] involved in supervised learning. It's really dynamic.

[00:06:07] You have these huge amounts of hyperparameters and all these

[00:06:11] interactions between different components of the pipeline.

[00:06:14] So it's a different kind of challenge and we're trying

[00:06:17] to tackle it from the auto ML side, but then also bring

[00:06:20] some RL knowledge in there.

[00:06:22] On one hand, it's incredibly like sort of like cutting

[00:06:25] edge and sort of like on the frontier.

[00:06:27] On the other hand, as with so much research,

[00:06:30] it's speculative and risky, right?

[00:06:32] Like is you never really know what's going to like transpire

[00:06:35] in the next like three, four, five years.

[00:06:38] Do you imagine that like reinforcement learning is going

[00:06:40] to just look pretty different in let's say five years

[00:06:44] thanks to its, it's like sort of like a conversation with

[00:06:48] auto ML.

[00:06:50] I mean, I hope so.

[00:06:52] But I also hope that auto ML might look different because

[00:06:56] of what we can do in RL.

[00:06:59] I really hope it goes both ways because since we have this

[00:07:04] let's call it more involved, I need dynamic process

[00:07:08] and we're really kind of forced to look more at paradigms

[00:07:12] that do for example, dynamic hyperparameter optimization.

[00:07:15] So that's a strong motivator I think also for auto ML

[00:07:18] methods to look more in that direction, kind of

[00:07:21] think more out of the box.

[00:07:23] And I think that could really be a nicely positive spiral.

[00:07:27] There and I think there's a lot of, a lot of desire to

[00:07:32] really make RL work better out of the box.

[00:07:34] And I think that will also help RL research tremendously.

[00:07:38] So I think the community is ready for that,

[00:07:41] whether that's more on the side of auto ML

[00:07:43] for a lot of meta learning or to melt for optimization.

[00:07:45] I think it won't even matter, but just getting

[00:07:48] there where we can say we can tackle more interesting

[00:07:50] problems, we can do the research.

[00:07:52] We're doing easier and faster.

[00:07:54] That's just going to be really important I think.

[00:07:57] I can remember in the first auto ML conference there was

[00:08:01] not maybe not a track but like certainly a few

[00:08:04] lectures and folks coming to speak to us about

[00:08:07] reinforcement learning which I thought was very eye

[00:08:10] opening because on one hand it felt like the

[00:08:14] communities sort of evolved kind of like separately.

[00:08:18] But at the same time there's so much that feels like

[00:08:21] it can hold them in common.

[00:08:23] So, well I hope that with you hosting the show we'll

[00:08:27] be able to hear more from the RL folks who might not

[00:08:30] even be in the auto ML community too and just

[00:08:33] pull them in.

[00:08:35] I hope so.

[00:08:36] I really hope so because I do think there's a lot of

[00:08:38] actual autumn RL research going on in reinforcement

[00:08:41] learning and people don't realize it.

[00:08:43] So it can only be good for everyone to kind of bring

[00:08:46] things together in that way.

[00:08:48] Make sure we kind of speak a common language between

[00:08:51] the communities and push forward in both ways.

[00:08:54] I imagine you're not going to be doing a lot of

[00:08:57] self promotion when you're hosting the episodes.

[00:09:01] But now I want to put you on the spot a little bit

[00:09:05] because otherwise you might not put yourself there.

[00:09:09] Where can people find you?

[00:09:12] Are you on Twitter?

[00:09:14] How can people find you and follow your work?

[00:09:17] I am on Twitter.

[00:09:19] I'm also on LinkedIn though I guess I'm not

[00:09:22] tremendously active on both.

[00:09:24] But you can certainly contact me there.

[00:09:26] You can keep up with my activities.

[00:09:28] Something that I'm also involved in that I think

[00:09:31] is cool to promote is the Co-SEAL network.

[00:09:34] It's about meta algorithmic research in general.

[00:09:38] We have really cool workshops each year.

[00:09:40] I'm one of the three general chairs there.

[00:09:43] If you are interested in meta algorithmic research

[00:09:45] that can be a really cool place to see what other

[00:09:48] people are working on, take some inspiration.

[00:09:51] I'm also involved in the organization of the AutoML

[00:09:54] conference which I'm not sure when you're listening

[00:09:57] to this but if it's still before September 2024

[00:10:01] come join us in Paris.

[00:10:03] It's going to be great and afterwards there's always

[00:10:05] the next edition.

[00:10:07] I went to a Co-SEAL one last year in Paris.

[00:10:10] Could it be?

[00:10:12] It's there in Paris.

[00:10:14] Yes.

[00:10:16] Me too.

[00:10:18] We must not have crossed paths.

[00:10:20] That was in which university do you remember?

[00:10:23] Zabon.

[00:10:25] Yeah, Zabon that's right.

[00:10:27] Can't believe we didn't manage to actually meet them.

[00:10:29] But you know what?

[00:10:31] We're probably in the same group photo.

[00:10:33] Were you there for like the big group photo

[00:10:35] in the courtyard?

[00:10:37] Got to look you up now.

[00:10:39] That was an incredible time and I also want to plug in

[00:10:43] Co-SEAL.

[00:10:45] Let me tell you something about the poster sessions

[00:10:48] that we did there.

[00:10:50] When I came back to New York from Paris,

[00:10:53] I was so inspired by that idea of the poster session

[00:10:57] and the way that it played out that I had come

[00:11:00] up with a totally different idea for how to organize

[00:11:03] AI events like industry level AI events in New York

[00:11:06] and it's something that I call ML show and tell.

[00:11:09] And the idea is just like in one of the things

[00:11:12] that I really love about these poster sessions

[00:11:14] is you could just stand there and just like

[00:11:17] grill one of the researchers until the thing makes

[00:11:21] sense to you because people, I feel like people come

[00:11:24] into this space from such different background sometimes

[00:11:27] like AutoML itself.

[00:11:29] People arrive at it from totally different kind of paths

[00:11:33] and AI and industry probably even more so.

[00:11:36] People come at it from totally different origins.

[00:11:39] And so one thing that I've always seen is that

[00:11:42] when you sit down and just listen to a lecture,

[00:11:45] I think it's very likely that most of the people

[00:11:49] in the room don't actually understand what the person

[00:11:52] is talking about.

[00:11:54] And they just sort of sit there and nod their head

[00:11:56] and sometimes pretend that they understand and most of it

[00:11:58] just kind of like flies right over their head.

[00:12:01] And I feel like that's the case in a lot of like

[00:12:03] the seminar style events that I have been organizing in New York.

[00:12:06] So then I thought to myself, but that wasn't my experience

[00:12:09] at COSEAL and it's not because I understood the material

[00:12:12] ahead of time, it's because I managed to actually

[00:12:16] interact with the people there so deeply.

[00:12:19] So then I started this ML show and tell

[00:12:23] and the idea is we bring AI practitioners

[00:12:27] and engineers and ML engineers into the room

[00:12:30] and I give them like a tall table and a monitor

[00:12:34] and they need to showcase some demo

[00:12:37] of something that they've built,

[00:12:39] but be open to revealing the code

[00:12:41] and going very deep into the code.

[00:12:43] And so I put on a few of these events here

[00:12:47] and they've been massively successful.

[00:12:50] A lot of people really like them

[00:12:51] and now we're scaling it up.

[00:12:53] So we did one in San Francisco

[00:12:55] and a few places in Europe.

[00:12:57] And it was all thanks to COSEAL.

[00:12:59] So lots of inspiration coming from there.

[00:13:02] So thanks for all the work you guys are doing.

[00:13:05] Yeah, if that isn't a great plug, I don't know what is.

[00:13:10] Teresa, the community is excited to have you

[00:13:13] and thank you very much for doing this.

[00:13:16] Yeah, I'm excited.

[00:13:18] I'm excited to get to know a lot more

[00:13:22] other ML researchers and a lot more other ML research.

[00:13:25] Let's do this. Let's dive in.

[00:13:32] Oh, I forgot to record.

[00:13:34] No, no, no, you can see it.

[00:13:35] I saw the recording.