[00:00:05]
>> Ok Welcome to m.l.@ g.t. lightning talk of we have a chock full day we have $22.00 labs showcase newday the lab will give presentations to 3 minutes this is only a small sliver of the research that goes on in the milky m.l.g. the image and Elegy was a interdisciplinary research center founded in 2016 and our mission is really to create a community of researchers that is very much in the in the did interdisciplinary spirit that Georgia Tech.

[00:00:40]
You know we have a tremendous range of research and trust in the center you know what we'll hear about today are things including robotics and computer vision and natural language processing privacy and Ai computational finance and even much more so the point for today is or basically the presenters to share what they're doing turn the research.

[00:01:03]
I did that and they you know the point is simple so one is to find collaborators with other faculty and maybe find graduate students who are looking for advisors or interesting research process Ok so this is going to work in a highly structured way we're going to give your exactly 3 minutes.

[00:01:18]
And we will start with what you've done or my slightly high sense thank you for the introduction just an empty happy to be here and I think he talked about that he so much and I now have. They'll am I research group is called fathom although it's it's an Indonesian we're not going it outside of the group but you know I have 4 amazing students and we work on a lot of problems bridging the screen in country as a.

[00:01:47]
So the f. in the Fathom starts what Pastor are fair algorithms and and we used techniques in continuous and discrete optimization to actually bring get faster algorithms reading in applications from engineering to means just from publicists they want to minimize non-convex objectives over some structural quality don't. Move on to the next time mission this end.

[00:02:11]
One of the areas that we are predominantly working in is the feet directions which are very I think an accountable and transparent decisions and we look at you know various problems in operations research in our computer science and we look at what happens when you look at by state and how can you actually make decisions that a fair across various opposition groups we also think about these operations in terms of fairness in service for example in electricity distribution networks can you have a pessimist in terms of the struction Xin the network and have minimum be connection times across different demographic groups the research like that I want to talk about briefly today is a quantum computer vision and benchmarking that we've got in recent interested in and this is a project with diaper.

[00:02:59]
Trying to. Bring in take me from funniest documentation into quite a militants and speed them up city go on the next flight you will see a quick overview of you know to get a flavor of the research that they're working on so if you're trying to solve Max Skype which is one of the most famous and had problems in control documentation with the quantum approximate optimization island and if you think about this algorithm this is really trying to get to a specific rotations of.

[00:03:30]
The quantum state where the state starts with a 0 which is an equal superposition of all the cuts in the in account and you want to visit the optimize these traditions are a Preterist and find the best one so that then then this data sample you actually get the max gotten a pass and the question we started thinking about is in one starts improve or degrade the performance of this algorithm so once taxes a technique that he is writing is optimization are.

[00:03:57]
Very. Very vital and the question was you know how do you even initialize Hewitt with such form starts and what we did was we took the when trying to solution scrum pyramid or not. Stations Golding Mark Scott in a 2 or 3 dimensions piece in the maps those are the current quantum representation and if you see the next animations this initialization of the quantum argue with him had very nice properties in terms of flattening out this piece in the ng it easier to optimize as well as improvements in the performance if you go into the next animation thank you improvements in performance you see though the blue regions out the van and actually I do it him and the red in the other regions and what we can do when we bring in tactical techniques into this though very exciting Seto for problems in quantum computing and a lot of a partially cook libation across you know their different.

[00:04:52]
Research Group is working on it and I think I'm over it and thanks so much Chris Lawrence thanks a lot without having it's gone to yes she another professor from Iceland. Laurie one thanks for an introduction Mike With pleasure to be here introducing our labs Ok so my researchers mostly git data driven and then Jim missionary because I find a lot of inspiration from various data that speech please Ok great so I summarize them as they go that are generated by physical sensors as physical sensors defense of your physical sensors water quality monitoring sensors and how sensors to monitor how status and he says are had generated a question I real data collection of measurements over space and time some are not for x. and also human sensors if I think about how we as losers of the system everywhere generating lots of data over time attention only or intentionally such as social net where it's electronic has records transactions and police reports medical records so we can understand aloud about the dynamic of this system the status of the system and see what's going on so these are different data generated by their sensors next page hooves.

[00:06:13]
And so in particular which is a funny signals or interesting from these noisy data it's a very difficult task because data can be very large and so it's really a finding a needle in a hay set and because the else is very complicated so some examples have we have been working our including the attack in the Middle East and now with me such as traffic incidents this and predicting whether or not or how long this is going to last and so it's threats for example a chance to talk to and that's me tremors and major events in social that's risky and we detect there is an outbreak of the disease say some people search parties and find natural collection is there any correlation or a calamity that the young correlation we have been working on for example look at police records finding quite link to just fighting called will not work in the brain not 1st looking at the neuro psych data they can predictions can't begin to future events in the past this is a big topic and how to what extent can we do this can they explain this from the data to sickle and its mission to rip prospective and censor prognostic send they can with medical recantations and so on next pay freeze and so here's an example of how the different types of all reasons and medicines and models we have the using for softness problems including particular sequester changes sections to school methods as summations hypothesis testing media mass learning and with your learning and using self and much exciting point prophecies are not forced to capture these different linkages and correlations and particular for example we have developed the system used by the length of 3 subpar minutes to find the correlations or police reports and recently we've been looking at some of the Colleen I think that if you understand a thing that makes all right I guess I'm ready on time so if we're to ask me one question thank you that's why I don't like about.

[00:08:07]
The mentally shaming. Takes place priest. And he my left and next if he my left where using a statistical and a computational all age to look at a problem set of interest to us and this could be our Syria who problems could be education problems in particular the type of the statistical tools we have proper you're using will be things from competition a statistic time serious model and the work we are likely to do is for example develop guarantees all summer implementable all reasons and for a reason the work which is available archive is a new homotopic approach that we are thought to to our existing statistical problem and we also interested in problems which come from the deep learning such as a which kind of architecture in the neural network is are going to be a fair Rabbo in both Siri in a practice we're also looking to some recent progress along the line of for example stock has agreed in the sand which is typically used to train on your network and also this emerged a series such as your attention kernel listening few views those approach to see that what what are the extensions we can do with those kind of for it is it develop to true or.

[00:09:29]
I am the director of further into discipline or research on the foundation of the does science institute at a church attack for the interest of time I probably cannot the puck too much about it but I provided the your erroneous lights which is a triad thought technology you think chesty can find more info and there and one more thing is a way also working on the sets from a different applications this including existing collaboration's other those in the pile step a working only electronic health record data and we also have existing collaboration's always industry working ways that they tout that is from the power grids and also recently are we start working or start up on something which is a really interesting is about automated the vehicle and I'll stop here thank you.

[00:10:21]
Thanks a lot of charming. We have more numerous faculty member Johnny who is joining us in January nice way you nice easy. Access then everyone can hear me. So I invited decided to talk a little bit today next light on foundational issues in modern day the science what I call inference algorithms and optimality next life so what this essentially involves is the union of 3 value Ridgefield statistics.

[00:10:53]
About confident predictions about actionable information that we can extract from data optimization how do we 15 smarter than how do we understand the evolution of optimization algorithms and information theory and actually understanding the optimality of that is the stages of that we might design on our data sets next life so this involves the way that the research is usually done by me and it is some sort of intersection of modeling theory and methods trying to use one to inform the other what I thought about today are 3 major cuts in learning from human feedback non-convex optimization in mathematical data science broadly and also reinforcement learning the next life learning from human feedback is a very useful interesting model and by the diamond data science whether that's in trying to design new ways to perform diagnosis on not on humans using humans to label machine learning these are sets of various.

[00:11:51]
You know downstream applications on understanding human behavior of instance you know watching demonstrators perform a task and trying to learn to control by those demonstrations just to give you a quick example next late one can actually take this thinking of thinking about modeling and methodology in conjunction to all read but there is elephant combat isn't fast and this gives you no more than theoretical support for navigating the bias variance tradeoffs in the stats which I just quickly said that.

[00:12:26]
The next last is in non-contact optimization this is something that pervades various things coming in statistical thing. Processing operations research training machine learning Martens today one interesting thing about these problems is that you're often working with them and them on samples of optimization problems. And what that means basically is that you want some sort of concrete but addiction's for how these would perform so here for instance I sketched best predictions for how that and the algorithms we do on going on including the processing problems and you can see that this quite a gap between such predictions and also you know unnatural things that happen in non-convex optimization for instance been in the in the 1st question of this late next leg but you can do is actually place these gaps you can actually think about these problems from a completely different perspective using high dimensional probability and then design better predictions on complex optimization the last thing I want to talk about is that in forcemeat learning.

[00:13:25]
Which is extremely pervasive now that from you know without a games and the drug test beds all the way to modern health and one of the important things that we like to do here is try and understand the behavior of reinforcement learning out of that in these environments and you can bring this perspective to this next late showing that some algorithms able to actually adapt their difficulties and problems but as others I know I don't want to say much more I just go and finish on the finals like you know.

[00:13:55]
And I also want to last one and just a quick plug for what classes might be useful and also a class that I'm teaching next semester Thanks Great thanks a lot of Ok So next up this polar challenge from our competition science engineering school. Thank you and yes know I'm in such a professor in the college computing and I work at the intersection of machine learning and so I say Jim and specifically on developing scalable methods interactive visualization and practical tools for the Cure interpret the boy.

[00:14:31]
And the essays that 3 something that we're working on one is on auditing Ai biasing where I would build tools and methods systems my fair share of this at the press one under the category of it's a our help with finding we call intersectional by example you've got machinery models build them how do you know what I wish to grip that's a gen gender and female plus I darker skin how would that how do we find out whether the machine and what about work well for that or not so that's still and also now we're working with at the end which you will hear later on extending that kind of idea to auditing that language model a lot of work as focusing on building tools at the slicing to help people understand what's happening in visualize a chanst solid horsefly from above and explain to us all all after that category our we call framework or reviewing the definition so I know a lot of you are interested and our building demanding model unit is my model complex machinery models often times we actually don't really know appling photos like cement and block and how to understand Ok you're dealing with machinery model how do I model which specific neurons which specific players are really helping with that kind of decision and also in the context of why when it doesn't working well with under attack so what it's really being exploited by the attackers and their recent work as an explainer is one pool where we are actually building it for us to do that is an educator's where we found that a lot of these models even though I use a lot in practice I also taught in the planning avocado but it's actually not that easy selects awnings I have recently seen this in explaining a tool which now is also now using some of.

[00:16:13]
Local morning classes. And finally we are interested in also how machine running on people running used in practice systems which means that it's dealing with potentially. An episode of a situation where they attack or it could be actual bad actors or it could actually be. Nature for example buildings how driving a car and you have opted to tactics sometimes it actually doesn't work really well in dart.

[00:16:45]
Scenario So how do we discover what is going wrong like what did how do we protect the missionary model and that is an ongoing drop or I've got a project to focus on building on the chance so we can learn more about what we do at our report age and he's at right microscope or cup but they doesn't like to name things out for myself.

[00:17:06]
A bank York Ohio and also I if you're interested and I do find one broadly I class this one thanks great thank you polo next the interactive computing we have more proud so market Dell interactive computing and I work in the entertainment intelligence lab so we like to make hard Ai machine learning problems fun by turning him into games and stories so I'm going to talk a little bit about a project that it's been working on for ever because it's a really hard problem is called automated story generation and it's the problem of getting a computer to make something up invent something to invent a story a fictional story from scratch ideally by just taking a whole bunch of data and figuring out how to.

[00:17:52]
Compose with it so go to the next like let me tell you a little bit about why you should care about stories besides the fact that it's super fun there's lots of things that we as humans do with stories they're central to the human experience though some of the things we can do.

[00:18:08]
Next. So we can entertain with stories I think that's pretty obvious we all watch Netflix we play computer games we can educate people we use vignettes in our classes to educate people go next. What you might not think about as we also use stories to plan so I can tell you what I think is going to happen next when we go into some sort of complicated activity or what I want to have next but I can also tell you things that can help you prepare so here's what happened last time I went to Disney World so if you ever get to go there you might be able to learn from my experiences next and we use stories human to human generally to bomb so can we also use Doris' to get computers to bond with us or for us to bond with intelligent agents So let's go to the next slide so lots of reasons why we want to tell stories telling stories is also very hard so some of the kind of the technical challenges that we get to pursue are next.

[00:19:05]
Well reading right you have to be able to read to process text or Kora training machine learning model school next writing obviously we have to compose we have to be able to write to have to learn how to read and write the next we need to know what it means even for an ai system to be creative What does it mean to create next planning so telling stories about what happens next so how do you figure out how to do longest range planning with computer systems and next and common sense reasoning so how do we make sense to team and readers so if we go to the next slide you'll see a little bit of an example of the sorts of things that we can do and we use these large scale neural networks now and they've gotten pretty good at producing text and here's some pretty readable text that you can just generate without thinking too hard do the next animation please but what's really happening here is the system has learned how to choose words based on the words is the source basically saying to tell the story I want to pick a board based on the words I've seen before and this works pretty well but you see that the there's a still a problem the story tends to drift over time and I'll kind of give you an example of what tends to happen as we run these models these predictive models a longer and longer so if we go to the next slide show you the exact same example again so click to the next animation here this is the exact same story.

[00:20:21]
But what happens next is kind of interesting a predictive system is going to tell us the next. One more time please. We take the existing text we run it through a very large on neural network and the next thing that comes out of this is something about wolves we're talking about dogs now we're talking about wolves what's going on here well the the the hidden layers of these neural nets are not saying anything about how we understand stories what we do is we come in and say well can we build mental models of what humans are and so I think.

[00:20:56]
I'm running out of time so if we just kind of flip through here really fast what should we know what humans understand about the stories and then how does the change in the story change our understanding the the story and if you go to the next animation you can see that there's this discrepancy that comes out right though this character that we thought was a dog is not Wolf that's a problem we need to x. out what was happening we need to change what we're planning on doing and do something else and last night I promise is now you can start to think about story generation as the search space through lots of different possible paths and click forward twice you can kind of see how they're good pals and that has We're trying to identify the good has that to be more readable to humans and what I've just shown you is a whole bunch of different really hard problems all trying to come together into something that is hopefully going to drive us it by making fundamental advances to artificial intelligence but also making things that are fun and interesting to read All right thanks I'm not super bowl Mark thank you so much Ok that's not also from come from propositional sign digit her Hosh So yes I'm I'm an assistant professor in the a c. school so.

[00:22:08]
Great yes so I at the site says I love the. One be there so I could apply it to their assignment to send data science a machine running with emphasis on Mequon sequences so a lot of public is. Through a lot of fine back up the kitchen such as it beyond the proper care and then how to fix them to read for example we have been I believe on the cool with that I'm going to response.

[00:22:34]
Use the plumbing approach just to pick us up on a bit of the Us also used by c.d.c. the eyes can see exactly the crime this could be a spoke with some don't challenge there are many other activities that you can see that. Is gone not got as mobility spread.

[00:22:51]
Be sure rising interventions then one using go. Gentlemen data and. And then also ospital accounting sections and I hold sort of other projects though so in all of these projects the idea is to use machine learning network algorithms and book science with modeling and simulation and domain expertise to do get some actionable insights and help the public and experts make decisions next like and yet so the big picture of the work in our group is really focused on using these things and very interdisciplinary so so so we try to meld approaches from back comics social sciences physics biology and obviously the traditions he has areas which is Curie and I got them systems and machining us mystics to create new competition approaches for these problems right so our focus has been really on on trying to advance the state of the art in these applications so so so that's why I use the I like to call it to our group on profit and social good because we really get about the applications which have which which can make a lot of difference in my trade and for example public health is a great example of far.

[00:24:01]
Off some of the kind of things we'd be trying to talk about next like. Another big project which we have been doing is based on urban computing that the Christian actual About 3 where again we build a 2 based on our machine learning a data center that works to really help them understand how failures my cascade across different critical infrastructures systems so for example our tool or Burnett has been ever licensed feels ecological utility and then we try to understand how these heterogeneous networks are connected and how can we identify could you wonder below traits will be used a lot of people in approaches suggested optimization.

[00:24:39]
And other machine learning and algorithmic approaches to do our job there are the readers of interest too in my lab we have one way glass missions but the companies and security so so these are also some interest because a lot of them can benefit from the cultural advances in books but just the big Thank You to my lab you can see the members there are actively hiring the board to be as distant supposedly a lot of open grants so if you're interested free to contact me and you can check by web site as well thanks a lot great thank you to you Ok Next up from d.c. we have just on a regime.

[00:25:18]
Thanks Justin can hear me Yes All right thanks Justin Next slide please the medical center in my lab is on it because a lady is like well it but clearly because we were consistent visual engineering appliances the Major. Work we do is really this interaction between data signals multimode everything and the networks and then your networks and I want to tell you the kind of between the 2 What have how they really interact with each other and that's really the code of what we do and the man and machine little problems with the Atlas thing as you have been editing for a number of years robustness active learning specially when we have chips and that in the medical field for example in a meeting need to have some some some input from from the end user and then really at large we have a pretty very focused on improving performance to say but really trying to understand what is happening provided it's been if an interpretation of what the new unit will be doing and across the board because in education the cases when we have limited labels what do we do and do in those cases but we can provide to the interviews.

[00:26:34]
And the theme is the reputation I will talk about. Lightly So one of the middle of the cases we have been focusing on for for quite You years as a trauma is because driving and robustness active learning in that field and we have been really partnering with couple of companies and they just need to come up with this robust distribution that the Coke machine learning side of it.

[00:27:01]
And there are tons of different projects and we have ready to prosecute on and idea and here is we had like 3 kind of make the network 11 attempts to decide where in the real world the want of the next. The 2nd application is the health care we have been focusing the last 4 years on bringing examiner to Knology to the patient to the.

[00:27:32]
General traditions we came up with this device this universal with Emory 2 years ago where really being the eye exam to the patient is going to need to go back to what I was your life with Greg it will not last more than 30 minutes. And the back end of that is going to be anything that he does quite a bit.

[00:27:52]
Of the intelligence in the system and we have to Caton license and proceed already on this technology but the end of the day we are any providing the politest the tools to understand what the new network is doing at the 10 time to really have an active learning model the feedback into the whole system from a position all the way to the back end with the next and last line and this is and family in here this is sort of your physics interaction between the physics of the signals and then you are literally so this is a driven physics driven deep network that's what we have been doing and thank you thank you Alan and so can I shot at Ok the stock also from seed we have more.

[00:28:39]
Ok thanks Justin. So like everyone here I guess I work in machine learning the lens that I come at this is through thinking a lot about applications and signal processing and high dimensional statistics broadly speaking I'm mostly interested in thinking about how we can efficiently sample some complex phenomenon to kind of extract as much information as possible given some budgets so one manifestation of this idea is to really look at any machine learning problem we're trying to learn from some experts and ask how can we design the queries we're posing to this expert so that we can learn really as much as we can from as few queries possible and this is basically the prototypical problem an active learning but also comes up in and supervised learning transfer learning reinforcement learning something bandits and some of my students been working in all these areas the specific labor of this that I'm most interested in is when the phenomenon we're trying to learn about based on the surface is very high dimensional so maybe we're trying to learn a classifier and in a high dimensional space or maybe we're trying to learn a ranking or ratings a very large number of items where naïve really it would seem impossible to learn without a very large number of queries but what often happens is that there's some low dimensional structure that we can exploit and it makes it possible to learn using many fewer So for example we might be trying to learn from the classifier for learning classify for high dimensional data where the data is sparse or low rank or lies on a low dimensional manifold.

[00:30:25]
Or maybe we're trying to learn a ranking of a large number of items where the person's a person's preferences depend on on only a small number of intrinsic parameters. So in settings like this we can think about how to take advantage of this kind of structure to either design our inquiry possible to create a query process.

[00:30:48]
As efficient as possible or alternatively maybe the queries are just given to us we don't get to choose what observations we have and we just want to understand how to exploit a low dimensional structure to get improved theoretical results or to develop better inference algorithms and either case what this usually operationally boils down to is thinking about analyzing and solving some kind of optimization problem that this comes up a lot in in a lot of machine learning problems but I also work on a lot of classical signal processing and to fix problems where the same things come up so I think about high dimensional time series data or where we have a large ensemble of signals like in a sensor network or an array of antennas and we have the same kinds of issues there where we have some low dimensional structure in the data and we have to think carefully about how to sample things and how to efficiently extract whatever information we're after.

[00:31:47]
That stuff any of this sounds interesting to you I'd say drop by my office and Botha but maybe e-mails a better idea x.. Thank you Mark almost a our next speaker is also from easy Morse Code And. Hello everyone so my name is Morris Cohen and I'll tell you a little bit about what the migrant does we are kind of at the intersection of.

[00:32:10]
Machine learning and geoscience are our geophysics not the case Earth in space physics some of what we do is not a ceiling related but a fair amount these days is so appealing about that next slide please. Ok so you know ls then for low frequencies and we're all about radio waves in this frequency band the span private or who was a fireman to kill or hurt I want to say that the world is our research lab so we study our kinds of things including lightning which I love giving lightning cause I'm going to actually let.

[00:32:42]
Me study lightning which is a global phenomenon and it made these really ways we can pick up from 1000 miles away we can feel Ok lightning and we can remotely sense sort of upper atmosphere clearance or kind of really difficult to directly sense through other means but impact things like navigation and communications and satellite communications Well we're also very involved up a little more in a minute in a sweater size and if Imagine yourself we can do including like you know how you map things underground or you know a service protection with these low frequency cyber security the power grid or something and come up with machine learning opposition for us right now.

[00:33:23]
And let's go with Skip ahead to the next slide. This is one sort of engineering the application so the low frequency use one of the new things they do is a penetrate through metal which you can do with most radar and radio signals that you that you can think of so if you want to for example look at what's behind a wall or inside a metal box you would naturally want to go to low frequency.

[00:33:47]
Turned up with a pretty tough problem to mathematically through traditional signal processing needs. So you can sit there you can take lots of measurements behind this wall and kind of infer something about what's what's there but if you really want to do kind of like an inversion type approach then the classic method sort of break down and they're essentially not a bird it was a machine learning is one of the techniques that we're using to kind of do behind the wall and inside a box type imaging using low frequencies as our sort of illuminating wave so that and this is this is a picture of some actual spirit that we did at a facility I'm just up the a.b.c. where the folks that I guess the Army research lab actually put some abject behind this wall and then tell us what they were and we did our best to try to him for what they are so that was a challenge problem so get the next fly.

[00:34:39]
I think our biggest worry is the machine very recently has been a space weather forecasting so maybe you know nothing about space weather but you probably should because it has the potential to kind of knock out power for example for a large swath of the country for weeks and months on end things like solar flares and coronal mass ejections Maybe you at least heard about if we get a really bad one we know that we have gotten these in the past and potentially cause some very detrimental effect the other problem that all worried about is loss of g.p.s. for example which if we lose yes there's all kinds of the vast rift things that would happen including no Internet no cell phones no financial transactions.

[00:35:19]
And all that is very reliant on kind of space weather not being horrible so. We're looking into machine learning as a way to piece together what the very very complex problem starting at the sun and working its way to what's called a solar wind and strengthening their magnetic field and the picture on the right is it's ridiculous to come get a diagram of the physics response to one of use solar phenomena and physicists have basically come about a far as they can yes to me as every one of the arrow tests wholly different assumptions and so you can piece together a single physics model the ceiling is very useful because it can help translate.

[00:35:55]
I guess different parts of the model for different assumptions are made and in doing so maybe we can piece together a much more broader whole of machine learning base pay for the work we're leading a very large effort to try to do this including 8 other institutions with different expertise on the side of this and our president machinery models are kind of being employed here.

[00:36:18]
Next slide so here's an example we're trying to our map to detect things like solar flares and maybe even predicts like when we're writing Well what happens is there's some interesting stuff to do there are many as a dangerous phenomenon skip to my last line. We work with a huge amount of data so that's a big challenge for us hundreds of terabytes per year we try to remove noise and separating us out as a lot of kind of the machine learning or kind of general signal processing process required so we kind of have all the classic problems we have difficult times a data set we have a regular samples we have that as it drop out reappear in dealing with our additive.

[00:36:56]
And quite a data loss of money travel all over the world are sites like extra cost up there and you can say thank you all. Ok is that the talk show or just college business Ok so they say now for something completely different. I work on ethics and policy related to m.l. as a quick way of thinking through all the stuff we've heard so far at least a lot of it.

[00:37:24]
There's a lot of How about bias in m.l. and there's a lot of critiques of machine learning in fact there's a deep fear times and I got a vision this partially from critiques from from my world where people claimed that they were filter bubbles and echo chambers but they're all blaming computer scientists and that algorithms are out of control and and have to be.

[00:37:45]
Well regulated and also arguments about if you just hand over this it'll all work this being code and so some of the 1st things I did was was this question of whether or not people really understood the intersection of platforms for information provision and the fact the ideas behind machine learning theory think ideas like exploring exploits have to be understood by the law policy side so I spent 2 years agoo go on the policy team and one of the things that I said when I was there and even interviewing was computer scientists are really not good at talking to non computer scientists and they have to be better otherwise deem isn't monsters get painted on and people will come up with really bizarre and bad regulations so it's another example that some folks may have heard of this notion of algorithmic transparency.

[00:38:38]
And the thing was I enjoy sitting between both worlds because a lot of my friends are computer scientists would would say things that at 1st I had no clue what they meant but but that in any non-trivial properties undecidable So why is why didn't anyone think about this and I said I don't know what you're saying and then I talked to a few other people it turned out a lot of folks are channelling the notion of this theorem called writes a theorem which some of you might know from the Course theory but then I thought about well but there's no way that all these amazing things we've been hearing about.

[00:39:10]
Just get threw up their hands and say we don't have any idea what the system does you work within the edges of that so I wrote a paper with colleague of mine pointing out that there is in fact really impressive ways that computer science can corral what comes out of it so in m.l. the fears where we heard today that people are working on versions of interpretability That's a massive step forward and that's the sort of thing that starts to create some interesting questions that depending on how one uses machine learning you're going to have areas where public concerns having to do with due process when people are taking Ai or m.l. systems off the shelf and trying to make decisions about who gets what I like to call social goodies that's a big question and if we don't have a way to understand what's being done and how it's being used people will say Stop don't use it which is a problem because as you also know and many of the people in this event know there's some tremendous bits that come out of this so the same things I like to talk about are I think machine learning is amazing as a diagnostic tool we can show what's happening you can analyze it right so even more is talk about which is way out there we can sort of go away to it can I start to understand how this thing works because then maybe I can fix it.

[00:40:23]
Sort of a little more concrete very excited to be working with Professor good to kick this off we're looking at the ways you can use an m.l. system to deal with bias in hiring because when you have a huge number of resumes coming in there might be a waste of pull them in this just tell me to shut up it's about time but Ok just to or to be to close this the things that we try and do when I when I work with people and if you're interested is on the active side if you've got a question where you're not sure how voting is being done how things are optimized in society this is where m.l. is huge because we can actually show problems but also fix them and with that I'll stop.

[00:41:04]
Thank you have a very interesting. We have those from our own from companies trying to mention a kite. So we are not the machine a new method is to study single us out so I just want to know do they live we want to study things ourselves and this is actually because.

[00:41:27]
Like you recently we have this technology and I was took at I sort of put a sequence in because from and this thing goes out as a lotion on and how this how do you know problems be on 70. 8 For example we can this 10 percent of how cells different c h for example or how stem cells if anything the different sound types and how enormous hours trying to convince our and our long time ambitious question the answer is how a single flashlights as they will not see your at out animal next night piece because these problems look Larry and they should but fortunately they have because you help us understand that out that these are actually looks like that people go kind of these are that to me usually the other lives many of my machine learning methods.

[00:42:16]
So the data is culled from the high frequency they got and. How it looks like it's basically a large matrix every row costs down every congressman to 18 and I read that a real number wasn't. In this data on the telling those learning of protein in this kind of a trick for us in this is very high to mission and secondly it's very nice and 30 it's very sparse next like piece and I'll ever.

[00:42:51]
Think it's such a large amount of data we can perform different accommodation not pass that onto these 2 these are 2 examples of what we can do in this data set on the left really show that how we can plaster the. South and from the taxes we can identify different cell populations and putting surely we can find we can discover new cell types.

[00:43:16]
Our body and on the right it shows that we can find the trajectories that this is basically like of the men we cannot find the scripts plasters of the scouts and here we can find the continuous gaps to between south and east all he has need that machine is actually rising our data is a very high dimension and even want to have a low damaging that input yonder to understand the relationship between between the cells next likely.

[00:43:50]
So we are not leaving the conceit of the cells in the context of both time and space so we know that during a longest time cells will change and still go differentiated I'm developed and also the axes they felt have neighbors and because to ban the very complex traction becoming stout and be a match.

[00:44:17]
The m a meant understanding how these these tales change over time and how the. Culling captivity to the next likely. Through the sound actually are collected at the same time unity but you are going to find a reseller you present the only state that has any reach you want to present the leader is this they want to sort them along the time we just want to be close to the time and next light peace and another another.

[00:44:51]
Kind of problem that we study about delving time easily look at the South in terms of the South decision processes the left the tree on the last group present how it wants out the devising to 200 to devise what stage we have to reconstruct this tree from a certain kind of data that we have to understand how you.

[00:45:14]
Are writing the next one please. Next likely yes these are the last night Candace rings ample of how we can use those they shouted at. The matrix that I showed earlier to understand better how the best out types are so here we are doing anything to please you know both halves of the During our get your honest and very real cell I think so here we develop they see the lower rank matrix.

[00:45:46]
The sentiment that. Approximation methods to identically to different parts of the so this is me and government continues to thank. Thank you. Ok Up next you're going to come back to interactive computer give. Me an assistant professor at a school in active computing is all. Our lab is called a social and a language technology acar name start.

[00:46:17]
We do research in the space of that's where languish processing and communication most of all science go in the peace based apparently we focus on leisure learning follow resource the task though you have limited at the top all with a week shouldn't it learn a masters we also are constantly stick a pattern or a ship so generation languish ways different style and the personalization we study buyers' the language we use and the ability systems to mitigate have biases in languish We also understand language ha language reflect our self and our society in the space of computational social science we care about a whole people using language to exchange social support or get information Eola house communities we study persuasion misinformation sick news and all that there are patterns that are cheaper but I language next pitch.

[00:47:12]
I'm going to quickly go through 3 work. From my life but this year so the 1st ally is a. Limit that they pass so we all know that and what she learned he has great success in a lot of task but real world scenarios are like you know the resource the language we don't have a lot of training data in this case Hollywood how with the ways to do the learning well though my students have proposed this technique called the text mix up so basically we try to do it how compensation to come up with a lot of argument to date how to facility the learning not only for test classification we have in our also working on other Pascal like a sequence a label any name to any recognition and also semantic Parsi Well we have limited Cheney and a tap next page.

[00:48:04]
Ok the 2nd the line what our lab is working on is called a stylistic attacks generation and a sunrise a ship on the left the saddo So this is our work this year so we studied just the projected buyers the language so the presupposition the framing buyers in language the example here is that on the left you have that you have used article had a line called John that he exposed as an agent of there was challenged the Couldn't we other magically turn it into a relatively neutral point of view such as making accused of accepting a proper donation from a real challenge so we.

[00:48:46]
Can call their vehicle their framework to do that to kind of neutralizing buyers the statement into our usual point of view on the right a part of this is a reason to work on summarizing informal conversations think about our current haka our random saliva conversations we have the model to kind of summarize conversations into a short summaries of our I'm now at the g.t. has a great blog on this work next page Ok the last allies that we also use all the techniques and mission a to study our society so for example we have work on the stand how people use our land House Community School support and information we also the ploy systems to better help the patient start answer peaches exchange support we study like a collective petition from a language perspective see during crisis events our language we use change how we refer to it the from the patterns different the locations different users blast the pitch.

[00:49:56]
Ok so the 2 step some have hazel we work on that production processing machinery for the resource to be task a stylistic test on ration we do research on kind of you know social science we have a lot of work in the space of his speech misinformation persuasion.

[00:50:13]
On the right is that you see that with a lot from messages about our lattice so we are looking for peace deal master and on a quest to our That so feel free to email me if you are interested very Thanks a lot the Ok So next we're going to go back to the show College of Business and hear from some of your child.

[00:50:36]
From every personal finance or select other business and also the faculty that Excel for the masses and $101.00 competitions and then which is a joint program at the school of industry engineering and School of Mathematics I will put that on the fight for a ca Sr Gen Than that of the tipping faculty 15 Master students in various aspects of finance so I think the nice thing I think I'm going last is many of the things that my colleagues in the machine learning group so far are thought of ought to have applications to finance because financial market dynamic and changing every aspect of mission learning and natural language processing that has been discussed so far has application and also provides a very that set of data to actually apply some of these models but example it might be like a gift learning model to do it fitting strategy but again at the same thing out of enforcement learning model but as we know.

[00:51:32]
It in many many had spent many traders were playing through the same thing it becomes a battle of a lot of them to navigate this not just what I got of them but also the data they should have also fitness and if this idea extremely important for example when somebody's a pleasure to create God then the client patient machine learning more does that then what kind of issues that they can direct from the social issues economic issues this policy issues but also on fitness and face in the pic ability of that important though any aspect or mission learning applicable and very interesting to our lab that same thing goes to natural language processing in fact is collaborating with the another one that there's a lot of data in finance and structured data which would be useful to look at greatest aspects of it there'd be a knowledge graph again just thinking about any aspect of m.l. and then I would be in applications the financial markets would be off interest would also sit still in session called a machine learning because that would be something which is that important and in general the good doctor and sees is that other area of interest in tech.

[00:52:36]
I plan to begin to protect that we will complete are going to step in because of law change or other things so any aspect of this anyone is less impressed by them collaborate from undergrad school masters to the existence to faculty would be happy to discuss and collaborate Thank you.

[00:52:56]
Thank you Ok I think we're going to try not to run again. So yeah I'm a little curious of in the same professor and so I have a group doing a range of work on the intersection of machine learning sensor processing and robotics and obviously this includes a lot deeper need.

[00:53:20]
And really focused on of 3 major thrust I would say although it's a relatively large group so there's other ones as well. So the 1st one that we're really interested in is pretty much anything beyond to provide learning in machine learning and deep learning so. Having very few amount of label data semi supervised learning where we have combinations of labeled a label data various forms of transfer learning and so on so there's a whole range of projects on this.

[00:53:57]
When's that differ a little bit is also continual learning which is how can we continuously learn about new things in the world. Dynamically and interact only by for example getting a few examples of new objects and then continuously learning about new things like that. The 2nd part is sort of joint perception and the so you're making so not just using the planning to understand things about the seams in objects but also using it for decision making so typically we start in the simulation environments that are photo realistic such as habitat or matter court if you have dozens of those for example generated by 3 d. scans of homes and then what we want to do is for example learn policies to navigate.

[00:54:46]
A star. Apartments where you may not have like continuous control so you just be moving from your point to your point so the 2nd thing that you heard about the joint perception of this is making so not just understanding the environment or scene but also actually making decisions and we're looking at for example navigation in simulated photo realistic environments and we're moving that through robotics where you have continuous controls as well so combining that with things like reinforcement learning and then finally we're looking at things like multi-modal and distributor perception so if you have groups of robots how can they jointly perceive the environment and for information and then collaboratively perform a task and so as you can see it kind of across the lab we do a range of different works that start from fundamental machine learning where we publish in machine learning conferences computer vision conferences as well for some of the computer vision aspects but then also apply them to robotics currently we're really starting to move back to actual robotics where we we see in there by the pandemic that will be continuing that once this thing is over looking at how do you actually apply this to real world.

[00:56:02]
That have robotic applications and so learning in simulation for example and adapting that to environment. So if you're interested in more please if we're to reach out and sorry about my internet issues. If Ok So next the art from I assure you we have generally I'm Jimmy and I join as white as far and Mohamed is on my lap is doing research in the area off statistical much so Marty to analyze and your imaging they have to improve the health off human brain next happy.

[00:56:42]
The why do we want to study the bridge next one. Next one because it's the most a complicated organ of human and the same time is very vulnerable to many different kinds of these and injuries next when as I've just to give you a few examples of the projects we are having the vast because the next one 1st one is really intense or are we study the most aggressive type cochlea lost so much that I've seen here is ours almost any hurt John McCain who actually die from it so what do we study to use much alarming to integrate your imaging and see that he can love actual art they have assets that we can develop more precise treatment for each individual another project next is the Alzheimer's disease and we focus on developing which is already are with them to integrate different kinds of near imaging they have for early detection that's one we also study you traumatic brain injury basically we want to integrate clean what they how within your imaging and either since they have in order to appreciate their trajectory off the recovery and how conditions decline for its individual after the injury and the last exam hall is the Parkinson's Disease.

[00:58:11]
Outweighs. Analyze the big data collected by people smartphone health asked and they were on the order to monitor their house condition over time and they would have imagined next life so what is our methodology next one. Y.l. I state have but p n for size I want to say is we emphasize the integration of different kinds off by magic or domain knowledge and how to develop new modeling frameworks that are with to integrate the knowledge and disappeared them Barny next one so to give you a concrete example this is an ongoing project precision medicine for green cancer the difficulty office cancer is what we call it into a humor hydrogenate see what it means is that different area different location of the tumor actually has very different molecular characteristics next.

[00:59:14]
So the consequence of this is that the treatment that is adapted for one area of the tumour may not work for the other area to overawe the treatment will fail next one so the research question we have is come we developed a methodology which we caught and knowledge infused the global local data fusion and the end goal of this methodology is to be able to estimate the Gryphon the molecular status across these 2 markets that we can do treatment of meditation and even put it to this methodology is that we like to integrate different sorts of they have new permissions such as biopsy data and also creating imaging as well as what we call the mentality it's basically the biomechanics of mottoes of P.T.'s simulators to simulate how tumor cells invade and grow throughout the bread next month.

[01:00:14]
The previous month there's another one. Ok there's one that's not missing but it's Ok so I would talk about the last one the last the slide It's example of the clean eco use of our methodology so basically we have a patient who has a print to Mur and after the surgery montra to the for location of the tumor and we want to now which area out will have tumor recurrence doing there are many other area that will apply or are with which generates the original mass for to molecular markers to catch a marker Also there are known to promote tumor girls and we want to predict which location of the tumor will recur and eventually the truth is that our map has a pretty good accuracy in the prediction So the indication of that is that in the future maybe more events a therapy can be developed to be attached to the 2 didn't know how to design a fanatic operation so that's all about my life.

[01:01:27]
Ok. So this new computer science we've come to be developed. I will spend most of my time in this life so what we do in my lab we bring together these 3 discipline plans must learning new science and network signs and I want to explain what is common between this 3 this means in our opinion even though this is highly debatable I guess.

[01:01:55]
If we want to so the I think you sell the little intelligence program we have to understand how the brain solves because the brain is the only instance of general intelligence we have in the universe so it makes a lot of sense to try to get ideas from how the brain actually.

[01:02:15]
Performs whether Ripper hordes so we are looking into a mirror science mostly to understand how the brain is structured as a network as a real neural network how the brain develops over time. Starting from you know an embryo all the way to an adult brain and how the brain functions how the network actually changes whenever we learn something in.

[01:02:46]
So these are ideas from from your science the tools that we used to study brains and brain at work are mostly coming from network science network science is a field that you can think of it as applied graph theory. Together with other things where you really look at the at the network and you try to understand it right out it's very serious modularity how it changes over time how it's function affects structure and so on and how does it all relates to muscle learning where very much connection is in our approach to too much of the things we think that.

[01:03:32]
The structure of the neural networks that we use. Can really learn a lot from the structure of our brain that works I should say that the structure of current neural network status you know convolution on the groups and the list the amps and all of that even though perhaps it was initially inspired by the brain it really has very little to do with range so we plan really want to kind of go beyond this very standard get pictures so very briefly I want to just so the next 3 slides they talk about some specific projects this is a that of a product on the lifelong learning machines together with says Ok that we have seen before and in that case the the goal is basically how to do.

[01:04:19]
Continual learning from unsupervised prints of data. The next project is. If you go to the next like this. Has to do with. Actually understanding from brain that where it's. Actually let me because I understand that I'm out of time that just skip to the next like. In the next slide we are.

[01:04:45]
Learning from the brain how the brain solves them with the sensory integration problem. That can be kind of directly applied in the use neural network of The Exorcist. So the more the sense of integration and I think this insistence that that's all the help things you. Are thankful are competent on top internet stuff from my assuage we have few but they have regular access.

[01:05:13]
Can you can you help. Here well. Hi everyone. I am appear to show and most of the work in my group is typical problems that spanned the use of control optimization algorithms are by no less today allowed focus on another book in a group under the present one trying to talk to her general purpose artificial intelligence and when he is taking a more holistic approach to what's actually enforcement learning is now one possible direction to attacking general purpose artificial intelligence and in contrast to supply the non-supervisory learning we have once once to solve a static problem such as the commision happened in reinforcement learning we want to solve a dynamic problem this immediate understand mine meant and the intent to interact with and write memes must a part time job or over $35.00 and usually the agent takes an action and the environment response to that action but you need to eat and that gives a reward to the intent and the goal is to maximize the long term to watch of the.

[01:06:19]
And such a paradigm can can be used to model why very few problems since I really do think that plans for the next slide we show the empirical successes of reinforcement learning in many different problems the 1st one is the us a good template by Deep Mind which is not practical which was able to be the best player in the game and open Ai and not profit built case Luketic arm which is just a single hand which.

[01:06:48]
Learned how to play how to solve Rubik's Cube and what essentially people are using the reinforcement learning to train robots in which all environments or even in real advice. And just last week deep minds us are for the. Algorithm or incrementally all the time but we can still have solved the 50 year old problem of cooking 42 so this is a good reinforcement learning is very successful and possible but the problem is in contrast to those incredible successes other understanding charity called understanding of the important learning is minimal.

[01:07:29]
The focus of our group as presented in the next light is on not bridging the gap between theory and practice so we were content because conditions have been present any other crime any viewpoint isn't fighting sample competency to find a technology to the question is how much data should I collect before I'm reasonably confident about finding the optimal policy and optimal solution to many of these problems and we do this start by we recently developed a unified framework a mathematical framework to study a wide class of been presented and we are working on using the.

[01:08:06]
Using this new point to develop new all the lessons that can be proved to be optimal and the improvement learning is a very vast area is it is a big idea and we use the. Other viewpoints and I think bonds to look at many different problems to give an example after critic McKesson policy good in math or back to exploit optimization methods to solve the present problem in the off policy one wants to decouple difficult Shann and learning and such problems arise and I think that has healthcare so in the next life.

[01:08:42]
I present in order to view of many different things we do in our group. I talk to only about one percent learning problem but I can say on the left hand signals are different tools that we use to cost more lean optimization dynamic programming and control theory are huge and important problems and other problems that they will Connery supplications load balancing the scheduling problems that I think of computing wireless networks but I think it's considered the next great thank you people Ok so I guess last but not least we have the director of an elegy of off or so is a researcher can get to Peter vision in competition talk or piece of introduce or punish the.

[01:09:26]
Thank you so much Justin 1st of all I wanted to get everybody see me you haven't seen very good. Actually Al you can get rid of that's why I think people know what I look like in fact the minute they can see me those be good and so you know just as Justin said I am the director of the machine learning center and I actually attended this talk just because I want to be able to hear what all of the wonderful amazing people of the center are working on you got a good sense of the diversity of the center my personal research is in the area of computer vision just to kind of situated besides being the director of the center working very closely with Justin Allie who's And behind the scenes on this effort right now I do research in areas of computer vision I do have about 7 ph d. students working with me in addition to that I also have an affiliation that Google or Apple a small research team working on various aspects of the work under puter vision that has the holes in technology of trying to create systems that can see so we can understand how humans see and understand scenes and use that kind of builds machines of various different forms that can also see and understand the environment there and there's a one of the ways machines is cameras and that is achieved by actually trying to understand the physics of cameras and how the light represent that and kind of generates an image that can then interpret it for various types of purposes of course that kind of means that we also work on things like competition for Tara feed everybody now knows that we all have a lot more cameras in our lives than used to and you have to understand this is a camera now on a phone which is a different set of physics than the cameras it be used to have before the competition for profit gets into the whole science of understanding how light is capture and then the Sturt and how you can use this up among other things to make pictures but also how to understand these days of course you know that using phones like this you can take pictures and nighttime where there is very low.

[01:11:26]
That it light and actually does all of the computation on board on a small processor not just be able to create beautiful pictures but also find things that you take pictures of that's the kind of stuff we worked on in the past our specifically when I work on this video analysis that is my team focuses on the question of what's happening in the video and how many improve the quality of that you hope various types of patients but also then asking the question is how can you create that Yes Now one thing I don't tell you is you know this is all deductive and because one of the biggest increase and sensors in the world has been cameras they're all the world around us cars have an arsenal to understand how so also navigate on roads and stuff they are actually using cameras to doing computer vision then similarly again as I said think about cameras everywhere and the other part of your house magine is cameras for a lot of data and when you start talking about video you're talking about a lot more and in fact a large amount of data on this planet right now is video data which has never been analyzed or interpreted so that's one of the other areas on my team also then concentrates on the area of what the focus computation of journalism is just really doesn't much have to do with images and videos but it does some of the other we're just trying to understand is how do you analyze and understand how information is shared gathered and distributed the areas that my team is now getting into more and more especially at Georgia Tech is the area of information quality that images and video these days are bullies concerned about after Musha of information the videos and the fakes and stuff like out there also interest in actually issues of how information like radio to share distributed and also verify you know making it read a file lots of different areas that my team works on an addition to this and stuff so if you have any interest in these types of topics these reach out to me the best way to get to me is e-mail and if you don't care back on the 1st cry again.

[01:13:24]
Justin thanks Justin believing this I know this this has been a great event I want to thank all our present thirst and Alan McFadden heard our show the Elegies communications representative for putting this together at the going on very well I had a pleasure hosting it and I learned a lot.

[01:13:42]
I hope you want to get more information about what was presented today you know what I will learn more about the people the research project for a purse of all of you can reach out to them they're easy everyone to define with a Google search but also the event here was reported and there will be available on the Elegy You Tube channel in the very near future Ok with that I'd like to thank everyone for coming and we will see you around.