159 - Dawn Finzi: From Vision Neuroscience to ML Engineering (Psychologist in the Wild Series)

0:00Welcome to Stanford Psychology Podcast, where leading psychologists come together to share their most recent work. I am Elizabeth, and for this week's episode, I'm very excited to share my conversation with Dr. Dawn Finzi as a part of our alumni and industry series. Dawn is currently a machine learning engineer on the perception team at Zoox, and received her PhD in our very own Department of Psychology at Stanford. During her PhD, she studied the functional organization in the human brain using various methods ranging from diffusion MRI to deep neural networks. Here's our conversation.

1:01Hi Dawn, thank you so much for joining us. I'm really excited to have you here as a part of our new series. And my first question for you today is, what was the thing that got you into science and what was that journey like? So I think I've always been interested in basically why the world is the way it is and also how other people perceive the same universe. I think I have this really corny memory of being essentially 11 years old in band practice, playing the flute terribly, and looking around at everyone and just thinking about whether they saw the same things I saw. and having this essentially existential moment about, you know, seeing and hearing and perceiving, knowing, of course, that they didn't think the same things I did, but what did it mean to actually be part of this world and to see it in a particular way?

2:01And I think that really kind of speaks to a curiosity I've always had about why the world is the way it is, you know, and how do we each individually perceive it? And I think, you know, there are a lot of different elements here. I had a mother who was an academic, so she was a biochemist who works at the NCI now. But I feel like she really fostered a sense of curiosity and questioning in me. And so I guess very early on, that's what got me into science, though I certainly had some teenage rebellion where I instead got very into philosophy and French literature and so on, but came back to science pretty quickly.

2:44You know, in undergrad, I really knew I wanted to study perception and to study experimental psychology. I spent a few years figuring out what exactly that looked like for me. So I knew I wanted to be in science. I knew I wanted to be an experimentalist, but like, what particular subfield did I care about? And I actually ended up coming back to what I originally found in undergrad, which was visual perception and just being fascinated by that. But for me, it was a bit of a little roundabout path because I did a master's, spent a couple of years as a research associate lab manager at UCSD, thinking maybe I wanted to do something a bit more around reward and motivation in a cognitive neuroscience lab that also did a lot of work around Parkinson's and deep brain stimulation.

3:35So I tried that out a little but eventually came back to visual perception uh just because yeah enjoyed immensely and find it very interesting and once you got to your phd program what was that like did your research interest go through any evolution during your phd and if so how did that change and what influenced your interest changing oh yeah they definitely changed as I went. I think when I first started at Stanford, and so I was in Colony Grills Factors Lab, as I know you are as well. And, you know, I really started thinking in my PhD initially as an experimentalist and really wanting to like think and learn about data.

4:24And so my first project with her was like very shaped by that. So it was, you know, using diffusion and fMRI, and it was about, you know, the different visual spatial computations and structural underpinnings for lateral versus ventral face selective areas. But at the same time, I was also getting exposure to a number of different subfields. So I was one of the mind, brain, computation, and technology trainees. So getting exposed to a lot of different computational approaches that way. And then also I'd been increasingly pursuing CS since my late undergrad. And by the time I was starting my PhD, the AlexNet CNN revolution was like five years in and it was in full swing and everyone was getting very excited about that.

5:13And I just found the idea that you could use these models to understand more about the brain and to probe the brain and try and, you know, learn different constraints or predict it to be fascinating. So for me, this was a bit of a pivot in terms of just knowing that I really, really cared about this type of research and wanting to shift a little bit more into that. This also dovetailed pretty well because by the time my first project wrapped up, which was like 2019-ish, I was already affiliated with the CS department and was finishing up the coursework for the PhD minor. and it was about, and I've been doing some like side projects around feature visualization or more traditional stuff about trying to predict primate visual responses using CNNs and so on.

6:02But there wasn't really an opportunity to do this kind of like single trial predictive work in humans. But around this time is when the natural scenes dataset came out or was about to come out. I know you've had Kendra Kay on your podcast before, but just to recap, this is like this really fantastic, massive human fMRI data set that was getting collected at the time by Kendra Kay and his colleagues. And one thing that was super, super exciting about this was that it was in a few subjects, there were eight subjects, but responses to absolutely thousands of images. So it was like 73 ,000 images, single trial responses.

6:43And this is like a perfect thing to learn a linear mapping to. So it like works really well for the kind of work that people were doing with neural networks as predictive models of the brain. But then it has the added benefit that not only is it human, because it's fMRI, it's such larger coverage than you would get with like primate experiments. So you could ask really broad scale structural questions about why the brain was organized the way it is. And like, for me, the question that was really interesting around this was like, why is the brain divided into streams and trying to understand more a little bit beyond the ventral stream, looking at lateral and dorsal.

7:23So all of this kind of came together around the time. And also very luckily, you know, Dan Yermans is one of the creators of this subfield down the hall. So it became pretty clear to me that like what I wanted to do was combine all of this, work with Dan, work with Kendrick and Colin E. And so I think that was essentially, I think of my PhD a little bit in like halves, that was essentially the second half of my PhD, really working more on those types of projects. So more traditional, like computational neuroscience and pretty much thing on network models of the brain. So I certainly took different projects and different subfields and directions as I went through my PhD.

8:10And I think that's one of the really cool things about a PhD, is that you're almost encouraged to do that, to figure out what's exciting for you and what you would particularly like. How did you tie those two questions together in your dissertation? I mean, I see the common thread. You're looking at different streams, but I'm curious because methods changed a lot. So I was just wondering what was that like? Totally. Okay. So I will say one thing about that, which is that I think for most PhD students, it's not clear how it's going to tie together until you're at the end. so I think if you'd ask me like I don't know four years in how it all tied together I would not have been able to tell you I think I had like ideas about how maybe it could but it didn't really fit together but then once you're like kind of doing a retrospective and looking back on it it becomes clear like you have your own personal motivation for why you did that wandering path and often there is like a theoretical thread that links it as well it was related to your personal of my realizations on this because I feel like it became clear to me how to tie it together around then, but it wasn't or wasn't immediately clear because like, as you said, methods are very different and the subject matter evolves too.

9:26But it's kind of almost a fun experiment-y type part of doing the writing of the dissertation to make yourself figure out what is the thread that connects them. So you did all of these really cool work in your PhD and ultimately you decided to go into industry. And I'm just curious what that process was like, why you decided to go into industry versus academia and how that journey played out. Yeah, definitely. So I think it was clear to me pretty early on in my PhD that at the very least, I was interested in industry. And I think there are a number of different levels why that was the case. So a couple examples are like one thing that really resonated with me is the time horizon.

10:22So in academia, particularly in the type of work we do, you're doing projects that span multiple years, so two to three years. And while you can really go in depth that way, and that can be very fulfilling, I think they were quite frustrating to me about that. And I knew that personally, I wanted to experiment with shorter time horizons. Like I wanted to be able to work on a project for three months, wrap it up, see the impact, and then move on to something else. um people are very different in that regard but i think i realized that i i wanted to be able to get a bit more breadth as opposed to depth on a particular project um and then another thing that i realized was that i love being part of a team and i found academia a little bit isolating so obviously this is a bit field dependent and you know you're in a lab so you are there are people you're creating with but it's not exactly the same as having like a group of like 10 or 20 people where you all have the exact same driving mission.

11:18And I really wanted that. Like I wanted collaborative projects. I wanted to be working consistently on a team on a shared goal. In academia, it is very much about like your own personal reputation and the papers you put out as a first author and then subsequently senior and how you build your lab and so on. And then And I also, and I haven't been as successful at this, but I think I wanted to tie my identity to work less. Like I feel that with academia, right, it's supposed to be a passion project. Like I think I had one professor tell me once that, you know, if you don't think about your PhD subject in the bath, you're not cut out for this or you're doing, you found the wrong subject.

12:06And I actually didn't want that. I wanted to be able to decouple my sense of self from what I was doing. I will say that is much easier said than done. And as another former Colony Girl Spectre Lab member, Taryn Data Scientist, so Yopold Rotsky once told me, sometimes the call is coming from inside the room, which is certainly the case for me. I think like I will always get excited and passionate and feel deeply tied to the things I work on. But that was certainly one of them I wanted to play around with. And then, of course, you know, there are the things that maybe sound a bit more trivial, but make big difference to your life, which is like being able to choose where you live and having a bit more predictability and flexibility and so on.

12:55So I think for all of those reasons, I did want to go into industry. I think I also felt like the type of research I was doing was starting to be done a bit more in industry. Even though I didn't stay in a research role, I felt like I didn't have to move so far from the things that I loved about being in the lab and doing research. So once you realize that you want to go into industry and you want to find a non-academic job, what was that process like going into the job market after your PhD? So going into the job market after the PhD, I think, was an interesting experience, but it was deliberate about.

13:38And I think that was helpful. So even before, you know, being at the end of my PhD and going into it, one thing I did that I like highly, highly recommend was take advantage of doing an internship during the PhD. And so I know this is something that is becoming more common, but for me, it was critical. Both because, so for some context, I like worked as an AI resident at Google X for five months around the middle of my PhD on some like neuroscience inspired AI and then also on some LLMs for source code type work. And for me, it was super important because it helped me actually like test some hypotheses about whether I would actually like to do these things outside of academia.

14:20And so I think like whether or not it confirms that you want to go to industry or whether or not it means that you decide you want to stay in academia, it's a super useful experience. And then also, I feel like it gave me credibility. I think, unfortunately, recruiters don't really understand academics as much as they could. I mean, obviously they do to a certain extent, but I think it helped showing that like I wasn't completely new to industry and to these types of applications. relatedly I actually think that one thing I was surprised by is the extent to which like people don't really understand our degree and I've even noticed this among colleagues now people hear that we got our PhD in psychology department and they think of us as like clinicians or that we've been working in abnormal psych you know don't realize that I've been working on neural networks since 2018 and that like our skills are very applicable and so on so one thing I was like really careful about in the job search, and I think I could even have done a lot more of, was trying to reframe the experience and the skills we do have into a language that kind of made sense.

15:29And I got some good advice on this. And I think I'm actually not that great on it, but it was something I really worked on because so much of what we do is so applicable, but doesn't always read as such. Like even things like in visual neuroscience, right? Like we talk a lot about perception and representations, like everything's about representations. And of course, there is some of that, like in that same language and industry. But you know, anthropic is going to call it mechanistic interpretability, etc. There are little examples like that about we are doing truly the same stuff and calling it something different.

16:05And learning where that was the case, I think was helpful. I think also, I just talked to a ton of people, I found that super useful, particularly ex-academics who had made the same transition had really good advice and were often very willing to talk and then I think I tried to be really deliberate about what I wanted to do like I knew I wanted to switch being going from being a researcher to like being an ML engineer scientist type role but I wanted to keep some things the same so I like wanted to I didn't want to work like with large language models. I wanted to work on something like vision perception related.

16:46And so it was useful for me to like to figure out what I wanted to keep constant and what I wanted to change as opposed to just like being open to anything really or any application in the space. But each experience here I think is very different and there are a lot of ways to get this right. So what do you do in your job now? I'm a machine learning engineer at Zoox, which is a robo taxi company. And so basically on a very simple level, what I'm trying to do is help the car see the world around it so that the downstream components can prediction planner and so on can decide what to do with that information.

17:27So when I first started at Zoox, I was working specifically on 3D detection. So taking in camera, LiDAR, radar, and turning that into bounding boxes in the world around us with, you know, yawn, class labels, and so on. I've worked on a number of different things since. I'm now working a bit more on BLMs and so on and some more generalist models. But broadly speaking, what I do is visual perception and then now also LiDAR, radar point clouds, etc. An understanding of the world just for cars to drive as opposed for humans to see. Now you work at Zoox. I was just wondering, was there anything you wish you knew earlier in the process or maybe even while you're a PhD student?

18:19That's an interesting one. I think there are definitely some things that were easier as I transitioned than I expected. And then also things I wish I knew. So maybe I'll say a little bit of both. I think so some of the some of the things were easier than I expected in terms of just adapting to being a software engineer. I mean, there were certainly some learning curves. As a PhD student, I was essentially using GitHub to collaborate with myself, which is very different than when you're actually at a company like Zoox where you're deploying code to run on vehicle, et cetera. So there are learning curves there.

18:55But in general, like much of what I do feels very similar to what we did as PhD students in the basics of like I'm training and analyzing models. That's the same. And then also one thing that was a nice realization is that we're actually really good at thinking about data and solving problems. And these are skills that like you learn in a PhD that are immediately useful. And that was very reassuring to find out. But then there are definitely things that I wish I'd known, some of which are pretty like, I think a lot of the skills you can pick up pretty quickly on the job. But maybe I guess in terms of how I wish I'd viewed my PhD, I think I wish I'd used it to like explore and grow and learn more skills even more than I did.

19:49Like I tried to be quite deliberate about doing that. And even so, I think it's easy to want to think of it as a job. And I think it's important to want to think of it as a job. You're not like a student in a traditional sense. You know, you're part of a lab, you're a part of a research machine, and it is important to think of it as a job. But then at the same time, like you do have this freedom to learn what you want. Like it is fundamentally about like personal growth and learning and development and pursuing the subjects that you're interested in. And I think sometimes I lost sight of that a little bit.

20:22And now that I like really don't control the direction, you have some control in industry, but it's not the same. I wish I'd taken more advantage of that. Yeah, I think that's one of them. I guess there are other like little things about industry and academia being different that were interesting to learn, but nothing too crazy. I mean, so I think they're less different in terms of hours than I expected and in terms of things being a passion project. I think I magically thought work-life balance would improve, but I think if you're doing something interesting, you're going to be excited about it and people expect you to be passionate about it and hopefully you're in a job where you want to be.

21:08meetings and time are handled differently that was one thing i think in industry people are a lot more like strict and deliberate about their time things are chunked in 30 minute chunks you're not going to have those like two three hour meetings with your advisor anymore yeah but some of it some of it was very similar like one thing i actually really like about being an ml engineer is that even though it's not a traditional science role i still feel like a scientist because when you're you know, tweaking loss functions or changing architecture and then like running a bunch of experiments, you're essentially being a scientist again.

21:46Like I'm changing some different parameters, changing some variables, running the experiment, and then like figuring out where the results are and plotting them and communicating my results. So that's actually, that part's been really reassuring to me about just how similar that is and how I can keep that part of my identity and what I enjoy. Yeah. Thank you so much. Okay. Maybe I'll ask a slightly different question. How do you think having a PhD makes you stand out in your job? Yeah, I think having a PhD makes you stand out in a couple of ways. So there are obvious ways that you might expect, like we can critically analyze research.

22:30So like when papers, a lot of what we do, right, is seeing papers that come out. So a lot of what I do in my job, for example, is seeing papers that come out and deciding, you know, what particular aspect of their technique or the code they released or whatever I would like to integrate and try and test and training the next set of models or whatever. And I think as a former PhD student, you really learn how to quickly, critically evaluate research and whether or not you think something's worthwhile. And so that's certainly helpful. I think also we learned to be pretty good communicators. I like the whole stage of the process, right?

23:10We learned to not only just analyze the data or train the model or whatever, but then to make great plots and visuals and make a talk about it and communicate it to people and convince the rest of the research community that's valuable. And while, you know, now I'm not convincing the rest of the research community that's valuable, it is super useful for communicating to the rest of the team and to the rest of the perception organization. Another thing, actually, that I think our PhD teaches us that I was surprised that other people don't have is, like, a healthy respect for data. You know, we are so used to just, like, deep diving into our data, understanding the facets of it understanding where it's gone wrong like i cannot tell you how many times we have been training models thinking that you know the late the data was labeled in a particular way or evaluating on a particular subset of data that actually wasn't capturing what we wanted it to capture and it was almost a no-brainer for me i think as someone with a phd to be like oh well what if it's the data why don't we like has anyone visualize this but not everyone immediately thinks that way so i think that is one thing that we learned that is that can be very useful um other things i think those are some main ones i actually think a phd teaches you a lot of skills that make you stand out comfortableness with uncertainty is another one, actually.

24:44I think because we spent five, six, six hundred years answering questions with no answer, we're kind of used to questions being formulated that way and are quite good at being like, OK, well, what do we actually need to pin down and formalize to solve this problem? Whereas I think sometimes it's easy to get overwhelmed or you don't actually see where the gaps are or where the question isn't clear. And I think that's something that starts to really come naturally as you've spent many years answering questions that didn't have answers. This was a really great talk. Thank you so much.

25:39thank you for listening if you could leave a review on spotify apple podcast or whichever platform you're listening this from we would really appreciate it this podcast has been a labor of love by several wonderful young folks here in the department. And we've been surprised by the ever-increasing reach the podcast has had. Help us make even more people excited about psych by leaving us a review or subscribing to our no-spam, all-fun, sub-stack at Stanford Psych Pod. Or shoot us an email with your thoughts or suggestions at stanfordpsychpodcasts at gmail.com. Thank you and have a wonderful psych day.