Session III: Dimensional Shift: Teaching Machines to See in 2D and 3D

hello welcome back take your seats and you virtual Folks at home you can also take your seats uh welcome back um to our next session and as you're as you're taking your seats I'll start talking the next session is very exciting dimensional shift teaching machines to see in 2D and 3D boy last session was very Weld and Woolly with the Fantastic questions that came in I

remind you on slido keep the slider going we're now at Caltech 3 session three our last session of the morning features one of caltech's newest faculty members Dr Georgia yuksari assistant professor of community of computing and mathematical sciences and electrical engineering yaksari earned her bachelor's degree at the National Technical University of Athens in Athens Greece where she worked with Petrus Maragos followed by a PhD from

UC Berkeley where she was advised by jitendra Malik from 2016 to 2022 she was a research scientist at Fair in 2019 Dr gyaksari was named one of 30 influential women advancing AI by the rework blog and in 2020 she was nominated for Venture beats women in AI Awards uh it and in the gist uh the objective of our artificial intelligence is to develop computational systems that

can operate at a level similar to human intellect and that's been a lot in the news recently chat GPT is one example of these systems capable of formulating and responding to intricate queries using natural language Georgia will delve into the topic of visual intelligence visual intelligence and the creation of computational models designed to interpret the world through images visual inputs with both two and three-dimensional representations

it's going to be an amazing talk so please give a warm hand to Dr Georgia yuksari [Applause] great thank you thank you so much for being here and welcome back to Caltech I hope you're enjoying your day um thank you for the wonderful intro my name is Georgia I just joined the Caltech faculty this January and I work on computer vision and specifically what I am

passionate about and what I want to achieve through my work is to teach machines to see just like we humans do and so I want to start my talk by pointing out how brilliant we are I don't know if you remember to say that to yourselves but we are we humans are the Pinnacle of intelligence okay this was not funny this is true we are um

I Envision um this is which is our ability to see um has certainly played a role toward that achievement very early on before we're able to do anything really useful uh we are starting to understand the world but just looking around and very quickly after that we are using our intelligence already to solve problems we start by solving simple problems like stacking the cubes and toys

into big towers and very quickly or hopefully quickly enough we are able to navigate this very complex world of ours uh like for example when we're driving around a big city and a busy City without causing uh catastrophes these are only a few examples of our visual intelligence and it's really tremendous that we are able to do this and the big question is can we actually

build machines that do exactly the same um this is the question that I'm trying to answer with my work and I've been working on this for now over uh 13 years and it's been exciting and a very challenging problem to try to attack so in this quest for building machine perception the first question that a lot of us researchers in this field wanted to answer is

whether we can build computational models that can recognize objects and images this is the very basic skill of recognition so this is a very easy task for you look at this image I am very certain that you've never been in this scene of this kitchen before however you have absolutely no difficulty of predicting what objects are in this in this scene so can we make machines

going to let this play because it's um it's indicative of what we can do today in computer vision so the model that you do the output that you just saw was produced by um this this model but me and my colleagues built that's called mask or CNN so Mass Garcia and those are neural network you can think about it as a complex function that takes in

a single image as input and it outputs the desired Target in this particular case it's objects their names and also their their localization information their silhouettes and it does so for any image you can input any image no matter what that image is and no matter how many objects there are in that image it will be able to localize them and also name them so Mass

considering was a solution to a problem that was puzzling us for years and achieving this was actually a significant Milestone to the point where it was is actually featured on computer vision textbooks as because of this great achievement and its model is so simple this is what we take pride in it's a simple model that can be extended to tackle more tasks such as human pose

estimation and it works so well that it works frame by frame um robustly so here you see outputs frame by frame of this video without any temporal smoothing the most the model also led to more advances in visual recognition for example uh Beyond object localization such as recognizing human object interactions where the goal is not to only detect objects but also the human and the action

that is being performed as a verb and again doing so in the wild for many action types and many interactions Massacre CNN as I said is a neural network and it's it learns its connectivity weights through processing images through learning from images the Genesis of neural Nets is actually attributed to neurologists cognitive psychologists and neuroscientists way before computer scientists like me got to adopt them and

work with them it started with hubel and weasel in 62 in the very famous cat experiment to Fukushima and caltexvarian John hopfield in the early 80s which then Yang lacun was inspired by later that decade and figured out how to learn how to update the connectivity the weights of these neural networks by learning from data now this combination of learning neural Nets from large data set

is is dominating the field of AI today um I am very certain that you've all heard of Chachi PT is nothing else but a big neural network over 500 billion parameters train on very large data text corpuses of over 1 trillion tokens tokens or words um it follows exactly the same Paradigm as as our computer vision models you take a neural network and you train in

on Data and I know that there is a lot of speculation and a lot of discussion around chat GPT and its abilities but one thing it's very good at it's composition so it takes two concepts that could be very far apart and it figures out how to merge them and elaborate on them so when I was making this presentation I was actually I spent a lot

of time having fun with Chaturbate so I asked it I asked to write a poem about the seminar day organized by the Caltech Alumni Association and the style of Edgar Allan Poe so fairly certain that Edgar Allan Poe has never written the poem about the the seminar day I'm fairly certain about this and Yeti have absolutely no problem spitting out a very large poem about this

let's read it just to say to see what it says so it says Once Upon Scholars gathered weak and weary that does not sound like Edgar Allan poet but he it has a sad um the subway that he had to write behind his poems in the halls of Fame Caltech or the wise and learn dwell okay and then a lot more so there was actually many

many lives I had to actually crop it because it was so long phenomenal this is great with all all of us compare scientists that were working with these models I work from the computer vision side but seeing a large language model like this perform this well was actually a fantastic achievement that we thought we could never see um in so such a short amount of time

and so excited by these capabilities of these large language models what a lot of researchers are now speculating is that well why don't we combine this model a large language model with vision models like Mass CNN so that now we can actually ask more questions about images to see what what do can we achieve in this space of intelligence and one such model is uh Flamingo

it's uh it's it's from deepmind from Google deepmind and combines a large language model like the one behind child chipiti with a vision model similar in designed to Massacre CNN and what it promises is that now you can ask any questions about the image and it will answer it which means that it has built this intelligence about the world through uh through seeing from images and

so um I want to show you this example because I want to make a point about this so let's take this example this is an example from of internet and so I asked flamingo to First tell me what this is very simple it says it's the departure board at an airport correct that is true but obviously our world is a lot more than that and what

this image shows it shows a lot more than that and what these models promise is intelligence so I wanted to see well can we do more okay so I asked which way is good for which way is gate 413 to the right well that's not quite true but maybe it was confused you know I want to give a benefit of a doubt so then I asked

which gates are straight ahead ABC the okay thank you for the alphabet but that's not really the answer so what I wanted to say here is that despite popular belief uh and what journalists out there say and I'm sure that you are reading some of these articles um these models are not quite that smart yet but what is certainly true is that we have made progress

in understanding our world through images what Massacre CNN produces that the fact that it can localize all of these objects so well but even this combination of of this Union of these models with large language model and understanding the scene types that is a milestone but it's a rudimentary skill if you want to claim that we're trying to push the front years of visual intelligence and

understand the world through images we need to do more than just understanding the presence of objects in the image and one important attribute that mascar standing which is my own work and also what all of these new fancy models that New York Times writes about constantly is that they actually are not able to capture one important aspect of the world the fact that it's three-dimensional I

don't know how many of you own a Tesla uh Tesla's very popular with us tech people we love them because they have this amazing software that's the autopilot that promises to be able to drive the car around so you're hoping that this autopilot that can drive your car has an understanding of the world in 3D if your autopilot only understood that there is a car somewhere

in its view but didn't know exactly where that car was that is a useless feature I want to be able to understand not just where objects are in my 2D grad but where they are in the real 3D World if we want to build autonomous system that can navigate the world the same way that we do and so well there is a lot of talk about

how to build this artificial general intelligence and how to push the frontiers of computational intelligence I argue that there's no such thing as intelligence if we can build machines that perceive in 3D and this is the focus of my work right now building off of the achievements that myself and my colleagues have achieved in the past years what I hope to do is build models of

perceived in 3D from just 2D images and of course this is a very challenging task because our world is so complex and diverse beautiful but complex which means it's going to be a challenging computational task to achieving so in this direction um I have I have dedicated several of my years and trying to find what are the best Solutions in this space and I'm going to

show you a few of these projects that I think are very promising so what you see here is the output of a computer vision model that takes images of input and produces and understands the objects in 3D space so this is the model that we call cuber CNN which was published a few months ago and I'll be presenting at cvpr this June and this is a

one model that can tackle all of these all of the scenes and all of these objects with a unified architecture so cuberson and is also a neural network and it inputs images and it outputs these 3D objects for any objects that might be present in the images something that I do not know beforehand my model needs to figure that out but unlike mass or CNN the

previous work that I showed you were masquers in and only detects objects in 2D it has a the view of a world as if it was a flat 2D world through the image grid cuber CNN is able to understand objects in 3D it predicts how far they are from the camera and also their size and dimensions in 3D space so this is a very exciting model

and seeing it work is also is has generated a lot of enthusiasm in particular about new applications that we can build in autonomous systems augmented and mixed reality so case in point we wanted to stretch test Cube or CNN to see how well it would work in any sort of video footage of our world and so what you see here is a video of a human

a person walking around the apartment wearing an augmented reality device so it's these nerdy looking glasses that the human is wearing is walking around their apartment so this is a a data set so cuberson was was was trained on single images it has not seen videos and a certainly not seen augmented reality captures but then we apply our cubers CNN on a video like this and

we're able to predict to to predict and also track the objects that are visible in the scene as the human walks through that scene and so this is the first time that we can do this and it has led to a lot of potential applications for augmented reality that we're very excited to build in order to build more assistive technology technology that can help us navigate

the world today however and pushing more in the space of 3D Cooper synonym does is oops is able to predict these objects in 3D but it represents these objects as cuboids and this is a crude approximation in an approximation that is that is okay and that is enough for a lot of applications that we want to build in the autonomous space navigation with robots does not

need more would not need more information that presumably being able to predict these crude objects through cuboids as objects will be certainly enough but what cuberson and of course is missing and what I'm trying to imply right now is that it's missing the fact that objects have 3D shapes so these are three These are 3D geometry that is that describes any objects around us and that

we would love to be able to predict in addition to just their attributes in 3D so now the question of course comes and is can we predict 3D object shapes from images this seems like a very hard task we are seeing 2D images and we're asked to predict detail 3D geometry and before diving into this and figuring out the Solutions in this space we had to

ask a few questions so the first question is how do we even represent 3D shapes 3D geometry computationally we humans don't have a representation of 3D geometry we just perceive it but our models are computation models need to be able to represent them computationally and then of course is how are we going to fuse and marry these 3D representations with learning and learning through with neural

networks and so to help answer these questions we developed this deep learning library which introduces efficient ways to represent 3D shapes such as 3D meshes and I want to show you this nice video I will never miss an opportunity to show cute dolphins that represent 3D meshes they're nothing else but 3D graphs but how to efficiently computationally store them and use them with deep learning is

unclear and there are many various choices out there and python 3D is is a tool in a library that will help accelerate 3D deep learning research with providing these solutions to researchers and in addition to 3D representations we also provide a very critical operation that Bridges the gap between 3D geometry and 2D images and that is the idea of rendering now rendering is a common operation

in graphics if you it it takes a 3D scene composed of the geometry the texture that's the appearance of the object and camera and it renders it to the image you see outputs of rendering every day when you're watching a movie or if you're playing video games you and your kids playing video games rendering is what you see constantly when you're watching these these videos however

rendering is not differentiable so I can go back from my image to to my 3D geometry but we're making it differentiable by solving some very interesting mathematical engineering puzzles to allow gradients to flow back from the image to 3D and so now with differentiable rendering in place and having computational 3D representations of our objects and geometry we are positioned to actually try to solve this problem

of understanding 3D geometry from images and this project with my wonderful student we we did exactly that we developed an algorithm or optimization approach that takes a few sparse views of an object and is able to Output the 3D Shape of that object the algorithm is very simple of course after you solve the puzzles that I just described so given that you have three view three

different views of an object and you start from an initial shape that's very far off from the true shape you use the the existing use to update the shape and predict the texture and from a prediction you actually compare it to the other view the the the additional view that you have and through the process of differentiable rendering and flowing gradients back you're able to update

the shape so that it conforms to all of the views of the object it's a very um this is a very big picture view of the approach but more details if you're interested I'm happy to answer um in our discussion and so this with this method you can you can reconstruct any objects of of any complexity like for example in this this example right here that

it's only using eight views to reconstruct that eight views is not a lot of views to understand geometry and the shape right here and so we've used this approach to help businesses with products so someone that wants to sell their guitar and has only four views we've we've lifted it to 3D so that users can better see it and look at it and presumably want to

get it and we've also reconstructed statues um in the wild from I believe this is somewhere close to the Bay Area but this approach was using many images four to eight to go to 3D so now the question is pushing more forward towards understanding our world in 3D space can we do this for only one image now this is actually a learning problem because we need

to predict a rich output the 3D geometry from impartial observations just a single 2D image which does which does not reveal the full geometry of the objects so what we do is that we follow um we resort to learning we train a large model and we train on large data but now our model has to be 3D aware so we have to design it so that

it can capture the 3D information from 2D images and our data is also 3D data and we have to process it accordingly in tune with deep learning so I will not cover how we do this but I want just want to show you some results on some objects that we did with that we applied our model to which are everyday objects for example this kid's backpack

where we can now reconstruct this geometry from just a single view that you see on the left um from captures from our our own very own iPhone of objects that we have this is a toy from my colleague Inspire if you're a Nintendo fan you probably like to see it I love spiral and we can reconstruct this geometry from just a single image and we can

either apply this to imaginary objects like objects generated by Dali Delhi is a generative AI approach to synthesizing images from only a text description so this is uh I we said a marshmallow in the shape of a cat or something like this it is a marshmallow the liver it's a marshmallow shape of a cat and now we can reconstruct it in 3D and a lot more

fun examples of that sort but the point that I'm trying to make is building generalizable computer vision models that work in 3D from just a single image and I'm nearing the end of my talk um and I hope that I gave you a good overview of the Journey of visual intelligence and where to where we're going with it next so my work is trying to push

the boundaries of perception perception in 2D and perception in 3D but I'm also interested in pushing perception at extrema current solution to computer vision including the ones that I mentioned here and the ones that are exciting the community outside of worldwide actually make an assumption that we can only learn about the world with hundreds of millions of training examples but we humans are an example of

how that's not true I can learn with asvi as one example and I can I'm able to generalize very well and navigate this role and only see from very few positives so how can we build systems that learn with that efficiency this is something that we are have not unlocked yet and might require a huge shift in how we're approaching learning maybe move away from neural

Nets and use different architectures different learning paradigms but it's an important it's important aspect of the problem to always have in mind our current Solutions are nowhere near the efficiency that we humans learn with and I'm also very excited to be at Caltech because this is a place where we can apply some of the results of these of these achievements to important problems I'm working with

a fantastic faculty here that are building robots we have collaborations with a cast trying to build robots for this world and also the moon uh where here perception is important if you want to send robots to the Moon we're not going to be able to collect training examples we're not going to send a robot first to collect a few images of the trainer model that's ridiculous

we need to have these models work out of the box in new terrains and new domains which means that we need to push perception even further and I'm also very interested in applying some of these models for the good of our world and especially applying them for Ecology where there's a huge need of applying Advanced perception models to understand conservation and other such attributes and aspects

of the changes that happen in our world right now that we cannot quite foresee but will be able to quantify and I'm happy to take any questions thank you oh fantastic thank you so much and thank you for your questions um is how to teach computers to see like humans the right question to be asking or should we be asking what is the best way to

see that's fantastic question the latter um so we are inspired by what we can do but our Solutions might not be exactly the solutions that we are deploying right now in our brain first of all we are not sure how we are learning and how we do see and how we do understand um so this is also a very active topic of research but I also

believe that the tools that we have right now in our computers are not quite in tune with how our brain is is wired and working so while we might not see the solutions that happen exactly the same and we will likely not we should be inspired by all the things that we can do well so that we know what we want to achieve so humans are

used as the Benchmark as is the light at the end of the tunnel but how we're going to get there is what we're here to do very interesting it sets up our next question good ones by Dean dogger if I've said your name correctly here or in the virtual world can the biology Dean are you here there thank you can the biology of other species such

as Dolphins echolocation and navigation of water help us teach computers to recognize the world yeah and absolutely and um and to some degree we are actually we are not just limited to passive sensors like RGB so first of all humans we have two eyes so we have some sort of stereo that we are able to perceive that even with one eye but how we are inspired

by how other species learn from a variety of ways from the kind of sensor that we're designing that we're using lidar is such an example all the way to algorithms and solutions uh and again an active topic of research that we're trying to be inspired by it gets a little metaphysical here um so I'm going to gang a couple of questions together um do you think

artificial intelligence could create a new dimension within our realm and one that was trending for a while was is the Multiverse real oh okay session um I am not a big fan of the conspiracy theorist quote unquote that I see around me that I see from journalists that are trying to create this sense of fear around AI I don't know if you saw today we're nowhere

near anything sort of quite dangerous happening yet these models have a very low bar of what they can achieve it's great that they can achieve it but it's not quite as threatening in my view yet um so I would say that I'm not sure if we are there yet I am I believe that our current Solutions will not get us to a place where we should

or anywhere should be feel should feel scared uh if anything else is the best time to feel optimistic about potentially solving for some important problems that can help with the problems that we currently have and how to solve them and now is the the Multiverse the metaverse like I'm not I think it was the Multiverse but you can take it any way you want I I

need a beer to talk about this how much seeing me we're going to switch from coffee to beer in the afternoon I love that absolutely um and Gene one more question from you which I think is is in in terms of intention how do you prevent unintentional cultural biases from accidentally biasing bias training of your neural net or machine Learning System yeah I mean bias is

a huge issue whenever you are learning from data you're running into um you you have this problem of am I biasing my model um so we want to make sure and when we're developing these models and we're collecting data sets that we are actually covering the distribution of the world and we are able to have a good representation of every everything and everyone around us but

there's another bias that comes in that is who are these researchers like me that are developing these algorithms so even that introduced a bias that it would be great to think about and and resolve and that is why we have these amazing efforts at Caltech where we're trying to take researchers from all around the world and bring them here and teach them so that we can

develop things together and then they can go back wherever they are and they can deploy these solutions to their specific problems so that's it's a bigger question that is something that I'm thinking about almost daily and how to work on it's going to be part of us as academics to make sure that we have the best guard rails in place to unbiased our models so this

question was low and then now it trended so um Bob schmeichel Bob are you here okay or is it in the virtual meta or here yes okay and and it's a Puzzler can computers discern between mannequins and humans yes wait a minute more more explain it so um they can if you if you provide the right training examples of uh what a mannequin is and what

a human is uh their computer vision models are fantastic or Discerning texture and it turns out that mannequins have actually much simpler texture than humans it's very easy to make of distinctions now if the mannequin industry decides that it's going to build very human-like mannequins then perhaps it's no longer going to be the case very interesting okay well and it's Caltech you're going to get this

so answer any way you wish okay what is your favorite AI joke oh God ah you can think about that I also need a beer to think about this hey I'm here we love that but I think the Edgar Allan Poe was itself yeah that's a brilliant enough okay so um do you anticipate this is a little more technical do you anticipate 3D computer vision to

make radar lidar obsolete for autonomous systems yeah um it's I would say they're complementary so any sensor right now that might so lidar and radar data the sensors allow you to have some sort of measurement of 3D um by pointing rays and and measuring when these rays are coming back um and of course as you can imagine these are not always reliable not always do not

always work very well so software goes hand in hand with Hardware so we're building models that can take advantage of sensors but can also reliably make predictions when these sensors are not available so I would say that the answer is both we want our lever sensors but we also want to build complementary software that can work even without that and there's always a question of energy

so I show you this AR headset right if you if we want to build assistive technology that can answer questions you can't pack it with a lot of Hardware first of all it's going to start burning so you don't want anything to burn near your face so you are you have a limit of what sensors you can you can actually pack and censor so that they're

safe for humans so going all the way to Advanced sensors might not always be be possible yeah I think on an eduard kotolink are you here Edward okay we're in the Multiverse uh what kind of processing load is being applied when the system is being asked to go from 2D to 3D what is the computational growth scene the computational wire growth uh so uh so there's

one what happens and how much computation is needed so what happens exactly so this is actually getting a little bit more technical um but in order to make a model happen that goes from 2D to 3D you need to start fusing ideas of signal processing and 3D geometry and figuring out how you can learn in both of these representations well and this is where it gets

it's a very interesting problem not only from a philosophical conceptual perspective to work on it but also from a mathematical and Engineering perspective how to facilitate that and how to make it happen and because now you are going from this 2D to this 3D space you are faced with this curse the curse of dimensionality well now your 3D space is vast so you need to make

you need your model to be able to learn how to make very good predictions while having all these infinite choices of possible predictions and to make this and to constrain this is where we use sophisticated approaches learning and learning from data to make it happen speak sort of what is the most this is another metaphysical once written what is the most difficult thing to render realistically

in 3d water clouds hair this is a great question and this is where there was this other question before about how humans Learners is not like when you're trying to represent the world in 3D you are deviating significantly significantly from how we humans learn and how we represent the reaction we kind of don't know how we understand 3D and yet we as computer scientists now are

faced with making having to make a decision I am not a fan of being religious so I like to have all of my options open and have various variety of 3D representations that have the pros and cons either by by representing the world as a continuous like smoke where nothing is rigid nothing is unpenetratable everything is penetratable but then it's a learning algorithm to figure out

the level or making explicit representations like meshes and point clouds and using them to represent the world it's an open question for of research and I'm happy to discuss more is the things that we discuss with students today how was the best choice but the best choice right now is to be aware of all of them and use them for for the application you want to

build well dialing back um let's see to chat GPT for instance I mean there's certainly these accounts of like the Edgar Allan Poe was Charming um we've seen sort of accounts where people get involved with the chat being GPT and ask too many questions and ask them are you evil and it says yes I am um do you think chat GPT is the face of the

future do I think Chaturbate is the Face-Off is the future so uh what I didn't present today is about 200 slides of what gbt can to you from things that are extremely simple that if I show you you would just not believe it um so I do think that there is there are some aspects that child CPT might be very useful for this compositionality which is

its strongest right now attribute is something that I think we'll we'll see uh May perhaps penetrate the world in one way or another in some Industries um but no it's not going to be the face of the reality because right now it's not trustworthy and it's not reliable it actually makes a lot of stuff up a girl in Poe is an example um and for a

lot of the things that that we humans care about we cannot use it if I write a paper and I write with chat CPT it's going to say a lot of give me a lot of citations of papers that really do not exist so to me it's actually quite useless right now but where this technology will take and how industry will try to build on it

I think is it remains to be seen um I think we'll we have time for maybe two more questions um one is many of your examples are man-made environments that kitchen reminded me of an Ikea kitchen I have to say um all of these objects have a lot of edges and cues regarding Direction size and orientation how do these algorithms work literally in the wild in

nature in nature yeah so in the world um so in the well means different things so um we I consider these results in the wall meaning that I don't actually constrain what kind of scene I don't tell the model I'm going to only show you kitchens right now and so that's what you need to say I just show it one image and it has to figure

out whether it's man-made or not that's certainly true and there were some plants in there but that's that's correct so the man-made world is a world where there is a lot of straight lines a lot of like orthogonal structures which do not quite exist in nature and forests and this is where um I this is where we are actually a little bit behind as in we

actually don't have enough representation of these type of scenes right now in our models a lot of the emphasis on computer vision applications are around certainly around man-made scenes but certainly the same Solutions can be applied to Nature if we had data that represented that domain last question and just noting that there's a shout out to Carver Mead um his work on artificial Vision with Misha

mahowwald back in the in the late 80s so that's the Caltech lineage um Anonymous asks may I buy you a beer just to answer those questions uh sure I'm happy to chat more you can also find me at my office anytime you're a medical Tech I'm happy to answer any hot questions and hot takes about the AI field certainly thank you so much Dr Georgia yaksari

thank you thank you okay and thanks to all of you here in Beckman and online for keeping seminar day uh Lively this morning I would say noonti dreary but that was at Edgar Allan Poe and he didn't go to Caltech so we're going to change it to noon Tai chiri as we break for lunch lunch is almost ready but just for a few minutes please stay

with us those in the auditorium everyone online because we're about to unveil the 2023 distinguished alumni award recipients and there's going to be oh so exciting and there's going to be a special six minute video premiere so now I'm really delighted to bring to Stage the gentleman who leads the CAA team responsible for bringing this day back to you in-person seminar day and making it possible

for those of you around the world I don't know just waving to around the world in the metaverse um so please give a warm hand for Ralph Amos the CEO of the Caltech Alumni Association thank you so much good morning everyone how are you doing great presentations right good well I am delighted to stand before you I am your alumni director and every day we wake

up and think about how do you best engage 26 000 graduates around the planet right so there's a lot of us out there and this notion these kind of signature days are a part of that work and there's a lot of beautiful things that are coming but it's my honor now to introduce the class of 2023 distinguished alumni award recipients and they are Dr Nedra angera

uh PhD in 82 electrical engineering a to Ned will Ramsey professor of electrical and systems engineering at the University of Pennsylvania Dr Karen Maples BS 76 biology obstetrician and gynecologist Kaiser Permanente Dr Eugene Myers via 75. mathematics director Max Planck Institute of molecular cell biology and genetics in Germany and finally Dr Ken Kenneth susslick BS 74 chemistry Martin T Schmidt professor of chemistry Emeritus University of

Illinois at Urbana-Champaign please join me in congratulating this year's class of distinguished alumni award recipients [Applause] if you didn't already know this is the highest of the Caltech stows upon our alumni and we do this annually since 1966 and we have a lot of these people out there really helping extend the brand of this great Institute as well as having impact on this planet so we're

very proud of their work and thank you to everyone who submitted a nomination over the last year or in the past keep it up this is how we find what our alumni are doing um so I guess maybe a question is how do you get to meet the future alumni that might end up on the screen so we're bringing in some New Traditions and we'd like

to take a moment to talk with you about it and show you a little about something that we're planning and I want you to invite you into the work that we're doing so imagine this that in two nights we're going to have several dinners private dinners at your homes maybe at a local restaurant with tables filled with techers all right table protectors is what we're calling

it is going to be a new tradition that we will carry here forward and it's something that really brings forth the community that the Alumni Association is here to exist and we thought it would be very cool to give you a little insight with that and then I'll have a little homework for those of us in the room and the many that are online to talk

about what this will look like so let it roll and here we go [Music] thank you [Laughter] [Music] I'm Jasmine Bryant I graduated in 1995 I was a chemistry major and I'm Chris Bryant I graduated in 1995 and I was a computer science major part of the reason why we got this house was to entertain family in the Caltech Community it made us family and it

is family to us and so just by natural extension an opportunity to have some family over for a dinner was an easy thing for us to say yes to hi I'm Steve shell I graduated in 2001 with an undergraduate degree in mechanical engineering my name is Amanda piapani and I'm currently a senior at Caltech studying computer science Tara Gomez I graduated from Caltech in 2011 with

a PhD in biology Matt hartshorn I graduated in 2007 with a bachelor's in physics I'm Kamal pulungan class of 2025 in electrical engineering I am Gene hartshorn and I graduated in 2011 Bachelor of Science in electrical engineering I'm Cesar bocanegra mechanical engineering class of 1995 and Fleming house my name is Iman oshik I'm an undergraduate studying chemical engineering on the biomolecular track and I am a

junior-ish but I'm matriculated in 2018. thank you all for coming please enjoy the meal we're so glad to have you thank you cheers cheers cheers I was looking forward to good food good company great conversation it's always something that's destined to happen when you're in a group with techers I love talking to undercuts um I brought this thing called Data match to Caltech you you guys

know about this yeah my friend this year she was like Oh no I got messed with my eyes just being around like like many people who are ambitious passionate and just love what they're doing it's just been a great experience to be a part of that Community oh you're at heliojet do you know Zan of course I know his name yeah Sam's in my book club

Zan is another Checker good friend of ours yeah now Steve was at e-solar oh yeah it's solar yeah oh you know Steve I didn't recognize you when we came in beard and Laser definitely definitely things are different but uh yeah overall I was like oh no that's Steve shell like I I used to work there are a lot of friends that I have who are lifelong

friends who are from Caltech much much more so than from high school or from from most other parts of my life that's definitely true we love the Celtic family that we have any 18-a so the first thermodynamic scores Professor was writing the book as we went so like each each week we would all get like a photocopy of the next chapter right unbeknownst to us we

were actually proofreading and editing the book and so occasionally we would get these these unsolvable problems throughout my career I've always had to face new problems and learn new skills and Catholic Education really created a great foundation for knowing and having the confidence that I can go pick up something new my junior year at Caltech I had to work at the athenium and I got to

they told me to to wait on a table when I got there and I saw who was sitting there at the table it was Stephen Hawking wow did you talk to him I poured some coffee and I was so nervous that I think I might have spilled some of it any opportunity I get to connect back to my Caltech years I will take that opportunity to

connect back there was one time I was coming to meet Chris at Microsoft I'm wearing a ruddock t-shirt right and he's not answering the phone and some guy walks up behind me and he's like radicals you remember reticence yeah he's like I'll let you in no questions like no questions literally being a Caltech an opened the door like it was super cool to talk to the

alums and see you know how Celtic has changed how Catholic has stayed the same something about it that felt very comfortable and very real as an Alum one of the things I've always had since graduating is Caltech remains to this day the hardest thing I ever did I remember being in graduate school and feeling like I was in this tunnel where there was never going to

be an exit I really felt you know just very very overwhelmed I really liked that nobody was like oh Caltech is nice and easy it'll be super fun they were like no it's hard but there are things that make it worth it like you're gonna come out of this feeling like Superman right like yeah not right away maybe but eventually you're gonna realize it is this

I don't know like rite of passage and then you and then you're in our club and definitely what helped me was having others you know come and remind me that even though you can't see it now and it might be miles before you see it you will get to that ending right you will get to that light just hear of all these Amazing Stories from like

seeing how much contact has changed throughout the years so incredible it's just fun to talk with people it's always fun seeing some familiar faces and meeting some new ones um especially with current students so I'm glad that that you joined us in addition to the alum So yeah thank you so much for dinner thank you for dinner I would say to another Alum who is considering

doing this or thinking about doing it I would say don't do it for Caltech do it for the students and your fellow alumni it's giving back yeah okay [Music] so that's what we hope you will do at your homes around not only Pasadena Los Angeles but throughout the world two nights in July we're working to gain host and we hope that each and every one of

you find a way to do it and really there are no rules you could do it at your home you might do it at a Pizza Hut it doesn't matter you're just going to get together with Tuckers and talk about life love and happiness all right so it's going to be a very cool thing breaking bread is very important in human culture with that I will

say we will help you to bring the people to your house while people sign up and you raise your hand and say I'll host Iowa host we hope everybody hosts and then we'll have a chance to bring people to your home near you in your region or it might be 10 of you it might be 25 it might be two it doesn't matter right Brad have

a good time so look for this you have a QR code in the program scan it you can sign up today and we're looking for 50 to 100 dinners around the world all right sounds like a good idea awesome please join us thank you so much and thank you to Sandra for the great work she's doing also with daa have a great evening grab some lunch

go see the campus and we'll see you back here at two bye-bye thank you