TWed Lightning Talks Fall 2019 (11 Dec 2019)

Key Takeaways

come on we cool glasses yes explain to coca so welcome to the all 2019-20 talks so tonight's very special because we get to get the opportunity to hear two minutes from everybody in the tennis world constellation who wants to speak about their projects which mostly will be featherless world grad students but hopefully some featherless world you are peas staff if you want to spend a couple minutes talking about what you're doing you can do that as well mostly this is about the grad students sharing we are recording we're not streaming live we haven't completely figured that out yet we sort of have streaming into WebEx recording and we will post the video of tonight's recording to our YouTube site just a couple of autistics we've got the the names up here in the order that people register thank you very much for registering we got a couple of open spaces for people in the room who can be coerced into speaking who haven't signed up please all you need to do is when we get to the end if you want to come up there put your name on there and then go to town our rules are really simple and painless two minutes on pretty much anything you want to talk about but it'll probably be about your research and notes no questions lots of cheering after everybody speaks give them your appreciation the way you want to get appreciation from yours and optionally I think we ought to require it talk with the Elton John glasses you'll get extra bonus points if you talk with the Elton John glasses don't not further ado it starts with severe and no email no get pulls no get pushes just listening paying attention okay okay okay there we go the bless it sounded laptops close it is severe hello everyone I'm severe so physicians or clinicians often find patients with unexpected outcomes like maybe they're gaining weight or losing weight maybe there's some symptoms they don't know why the patients are having these symptoms may be some other unexpected outcomes or adverse events so in order to explain these unexpected outcomes what do physicians do well they actually combine the various various forms of reasoning so they do abductive reasoning deductive reasoning inductive reasoning abstractions all these sorts of things so we want to create an AI system that's able to mimic how a physician diagnoses patients and the way we're doing this is by creating a hybrid reasoner so using Semantic Web technologies we can already do deductive reasoning so we're planning to build an abductive reasoning pretty much given a set of facts in the knowledge graphs we want to explain how they got there [Applause] the glasses here talk about humor no this is better I can't see the audience I'm gonna talk about some work I did for the semantics reading course over semester I was looking at mythology alignment and I found kind of an interesting case that I just thought I'd talk about so for one of the survey papers an author went out and took a different on a bunch of different ontology alignment algorithms and tested him against three data sets and for one special interesting case the algorithm did great in its own paper it did great on the first data set and it did horrible on the other two and it was kind of a shock because this is a well known algorithm and they went and the authors actually kind of responded to this and they said well it is it was a good experiment you just didn't generate the correct hyper parameters and that was the big problem here this was a great algorithm as well but there's a bunch a large number of hyper parameters to customize it to the data set that this new third party user of this algorithm just couldn't adapt to and so this was kind of an interesting use case for me because I was thinking yeah we wait how long and Oliver the neurons its accuracy in certain data sets but not how hard is it for the next guy to use and this kind of example highlighted that and so would that new kind of perspective I continued looking at oh look how hard are the other algorithms to actually and how many hyper parameters did to work with and yeah so I guess just as a warrant lesson learned as you are developing algorithms in the future keep an eye on that because somebody's got to use that Sunday and if you want people to use it we should try to make it as easy as possible nobody's going to know their algorithm as well as you know your so these here also let people know who you're a by yours Oh Debra's my father thank you yeah paper idea yeah hyper parameter is considered harmful [Music] yeah hi everyone I'm Shruthi I am a first year PhD student working with working with working with working with Professor McGinnis today today today I will be talking about a project that we just started explainable health assistance essentially what we essentially essentially essentially are essentially the objective that the objective that we plan to achieve is to build a system to provide personalized trustworthy and also provenance of an health information to health information this semester we conducted a literature review in the vast and also ever fast and fast and also past and also about evolving literature of explainable AI so essentially what we did is to kind of categorize the papers into the categories of expert systems cognitive assistants Semantic Web and also the Machine joining domain additionally additionally additionally we also recognized that it saturations and context demand demand demand demand different explanations so with this task for this task we needed to look for the Buddhist task we needed to look for the explanation types in like in in the fields of social sciences philosophy exceptions well besides the besides the computer science domain what all he found that um Robbie found that contribution could be a semantic could be a semantic understanding of the semantic understanding of the explanation space accounting for the different entities processes and also the actors involved with explanations where they're at right now actually we are writing a book chapter on the foundations and foundations and foundations and foundations and directions of explainable explainable knowledge enables systems and besides that will be also other developments next semester so today so next I'm Jason so I'm a PhD student working with Professor McGinnis so today I'm gonna share some of my research during the summer in China and this fall semester so what we are trying to do is the fresher answering tasks over knowledge graph so for large graph question answering the segment a passing based methods aims that generates logical form representation for for an improved question for example if people ask what awards have have been won by the executive producer of the Timmy time which is British in a TV series and image animation for children so in this kind of relatively complex question so there are two triple patterns in involved for the logical form of the input question so to answer this question we need to first generate the correct logical form for the passing of the question so what we are trying to do is to propose a neuron network to generate such logical forms in triple pattern formats to help the task of transferring over the knowledge graph and then we have conducted experiments on row data sets which contains the question and spoke orientation pair for in variation and then the result is quite competitive so for future work we want to further a bra a combination of not graph embedding and the kg q a tax we want to see whether the graph embedding can pair positive will grow in generating better and more accurate logical forms because currently the knowledge graph embedding methods usually in varying using the triple classification and link prediction tasks but we want to learn some they're not graph embedding specific for the kgk task so this is this my sharing the people don't realize is those looks worse okay it also makes it harder to see no my name is Peter I am a master student under dr. McGuiness and in her class this semester I did some work with a semantic web based system for crowdsourcing the detection of fake news so fake news is a very well known and widespread problem especially in today's political climate and there have been some attempts already to make technological systems well alleviate this problem however all of them either use machine learning which often looks more at surface level aspects of the text and does not actually check its factual content as well as machine learning can be a can be tricked easily like a lot of people here I know when did that talk a few months ago about how to trick machine learning or else they require expert users of the system to be checking which creates a bottleneck and neither of these options has any accountability for the system so what I am proposing instead is this system where ordinary non-expert users can use this browser extension to publish publications now the content of articles they read and they can say this is true this is false this is true because this other source agrees with it this is false because this other source disagrees with it they can also say now publications about what other users of the system are trustworthy which will create this trust network which will allow the system to show users explanations why an article they're reading is true or false and give them like some provenance information of where the system got that recommendation tailored to the users existing trusts okay so I am miner Gordon I am a director of research under dr. McGuiness and I've been working on a project in my own time to create an ontology to describe collaboration networks so these are what they sound like networks of collaborators the classic cases in scientific domains people who publish papers together or who are in France together there's also artistic collaboration I'm particularly interested in one domain which is collaboration between nonprofits so last year I work for a technology consulting company that's only worked with social impact nonprofits and that's actually a really big problem nonprofits are competitive within a community they're competing for grants and so on and the funders don't often know what is actually happening in the community so the idea was to create a system of metadata for describing what actually exists the resources people services data sharing systems different sectors that are being served to the nonprofit community so I'm using that to kind of address that problem there's no money in that so I'm doing of my own time there was no money and event either always come with passion projects I'm also creating web application to visualize those collaboration networks I'm particular and in the sort of historical science collaboration networks and I'm using the web application and all the tools to you pilot a bunch of things then that I do here software stacks scalar triple stores things like that that I gear that I then take to work so greater domain-specific language originally did that last year for building ontology is so using that as a way to reintegrate my work as well PhD student professor McGinnis as my adviser and so as much as I love going out getting pizza like we're having today if you want to be eating healthy generally what you want to be doing is cooking for yourself when you're cooking for self you're typically kind of going to be going around a few recipes that you know you like but maybe you need to be switching up those recipes a little bit once in a while maybe you don't have a particular ingredient this week maybe you're gonna be cooking for a friend about starting a food allergy maybe your doctor recently told you hey you need to cut down on the cards that you're eating in order to do these kinds of things might be helpful to have some sort of system that can generate potential substitutes for individual ingredients within recipes maybe even finding other recipes that are pretty similar to a recipe that you know you already like so the project that I've been working on with the heals project is to be using the food kg which is a food knowledge graph that another student made that includes a bunch of recipes the ingredients that are used in those recipes as well as links to nutritional information and like the food class hierarchy of those ingredients and I'm trying to find substitutes for ingredients based on the context that it's used so you want to be eating potatoes or a flower if it's like a mashed potato but if a potato is being used in a student maybe I mean a different kind of substitute for it and so yeah we're trying to combine this kind of semantic technology and applying it to substitute ability and similarity in the domain of foods and recipes all right cool thank you hi I'm Owen I'm working with the DSA project under Professor McGinnis I'm an undergraduate researcher here I'm gonna be talking about I guess what I've been working on for the past month and hopefully working on for the next semester I've been working on designing a domain-specific language for describing the policies that use in the dynamic structure net axis project basically we'd be describing a language that kind of abstracts away from the a bit like hard harder to type rdf that we use to describe it so instead of needing to like put a bunch of RDF types dis IT policies we discreetly just describing using like a single keyword such as policy and basically this would make it a lot easier for a spectrum managers to actually develop the policies add on to them edit them and so on if you take a look like with the text-based interface basically we'd be applying this in something known as a language server som this is something that Microsoft's developed as a way to basically do you perform like a bunch of syntax checking that doesn't depend on like an editor so you don't need to write an entire stack just for like one editor to work with and it would it would allow for users to be able to search up for previous URL your eyes that have been defined so that can perform like a lot of Auto completions I guess some X completions and checking all within like any editor use ready claim for and so yeah hi everybody my name is Ethan Ram Bakker I'm a code term student working with Professor McGinnis and I'll be talking about my master's project from this past semester which is implementing a triple store is backed by elastic search elastic search is a very scalable distributed database document database and basically we implemented an elastic search back-end with Jana our EF API which is used by similar projects here I'm soulfully this project you know if performant could actually be used in these other projects and we benchmarked it evaluated against two other databases from Apache Channel TV to used in remembering and persistent TV two databases benchmark that against the elastic search backend and found that the elastic search back-end is not quite as performant as the other two beckons and I should mention we used the Berlin Sparkle benchmarks which is a set of twelve Curry's that I'm running against the sparkline point and you know so we found that basically these queries performed or on the elastic search back end then on the TV to in the future ideally this would be benchmarked against a few other options like place graph or Neptune and also hopefully optimize a little bit further as we haven't really optimized quite like each the optimization limits yet and all right hello everybody my name is Kara Reedy and I'm a fifth year student working with Jim Hendler if you've been around for a while you probably heard this a few times but I work with an artificial life simulation that I've developed in Python the simulation consists of an environment with various objects and agents that inhabit this environment and can interact with those objects the goal of the simulation is for the population to evolve over time to learn to eat food and not rocks or each other and hopefully be fruitful and multiply and exist for many many generations sometimes they die out after like a thousand time steps and after a run at 50 times but generally speaking they're doing better than they used to which is good so these agents is decisions are made by their individual neural networks and these neural networks are created by their genetics usually in something like a genetic algorithm we use happily genetics which have one copy of the genetic material that these agents are diploid meaning that they have two copies of the genetic material not only does this work better for the specific simulation that I'm working in it also opens up some more possibilities for future experiments so one thing that I've been intending to do for a long time that this way with the simulation is involving more aspects of social behavior between the agents and so recently I've been looking at kinship discrimination or the ability to differentiate between family and our family and to treat them differently so in my current experiment I've introduced genetic disorders into the population using this cycle a genetic system and I'm looking at how different kinds and prevalence rates of genetic disorders affect the populations ability to evolve inbreeding avoidance while hopefully still mating with non family members and if I have time before after I hopefully get out of here I would also like to look at how migration between different populations could also affect this rate of kinship discrimination most grad students hope that their work gets carried on but so you have to take care of hers yes you do Matthews did you find them there yeah so okay fun story we had high school and kids this summer code in there too yeah so this was Matthews funniest personal level yeah hi I'm Peter student Peter Fox's daughter dr. Fox is student today I'm going to talk about emergency response decision make if the project we started about a year ago it was basically trying to understand how emergency teams the information they have weather and Hawk situation they're presented with and how they combine the knowledge of both these domains and tackle with the problems they're presented with started an out thing we started out thinking that it would be as easy as a plus B equals C but it wasn't we realize there's no path to a right answer when it comes to these emergency situations so we scrapped the first set of models which we had and it was not an easy bread search first kind of a thing so now we're moving on to design and improvisation where it's like convergence of planning and execution so now we are trying to understand what goes into the planning and execution phases when it's such a time bound situation that you're presented with and the simulation was ran into forms where you are actually interacting with this emanation as well as having a group discussion parallely so we have this huge data set that we are working with and trying to find the nuances in it and making sure that we get it right so hopefully by next semester we'll be able to present some solid results I'm Sam I'm a new software engineer with the Telus world I started about roughly three weeks ago I haven't really done your research here so I'm just kind of introducing myself I was a student here or PI for my undergrad I went to work for the state government for a couple years and I worked for the state legislature building open API is that provide legislative data or people to make their own apps with I thus far here at the Telus world I've been working on the DSA projects working on some and testing using this tool that I'm learning now called Cypress and you have been really diving into the Semantic Web learning about all these technologies and thus far I'm really excited to be learning all this stuff and looking forward to working with all of you number two everybody I'm Chris I'm a European under the DSA project for this semester I worked on creating a request builder to basically allow LGS internal users to test their own simulated radio requests against our ecology and so basically the request builder allows the user to vary between just like a number of parameters location transmit power frequency radio translate video typing over and be able to determine whether or not that particular radio quest is permitted or not so you can get like a permit permit with applications or permit with reasons as to why information doesn't go through and yeah and next semester I'll be working on a different project where with temporal and spatial alright go out with a bang there we go there may be more people that's true well we don't know um I have been up for about like 30 hours so I really bear with me my nerves will be down so I'm I'm Nick Kerry I'm a senior here physics student working with dr. Fox for my senior thesis and my senior thesis mostly focuses on trying to Bill actually I'm really doing this so I can talk to you after so it really trying to help understand model collaborative networks specifically pertaining to any educational space so when you're working with a variety of different types of students disseminating effective information actually getting them to communicate and work together and then also as subjecting them to experiences is very challenging right yeah yeah but probably all know from being TAS or professors or or whatever but when you're dealing especially in places like I'm looking at right now is in Ghana you have a really wide distribution of Education levels and what people are trying to do and are like understanding their understanding and perception and so at the same time I'm trying to explore is how we can build a type of collaborative network that can help individuals process and receive firm understanding how to complete projects and then be able to effectively work together to address complex challenges with who is in their communities so my research have mostly spent this semester been trying to understand just you know I come from physics so my understanding of network science has been limited I've been trying to understand more about network science and understanding some of the principles and now next semester I have been planning to go to Ghana in March to do a oh if I get the money to do it so we'll see to actually do a more ethnographic approach to trying to understand and collect qualitative and quantitative information to better oh great Network science came out of physics and yeah alright so people are dying in the United States at higher rates than in the rest of developing world we got to do something about it to quote Kristin Ben so a group of undergraduate students have been working with us since late very late in the spring 2019 term in developing a app called mortality minder which is an entry in the health and human services agency for Healthcare Research and quality social determinants of health visualisation challenge and this is an app so in mid-summer we won the first round we won $10,000 for being one of 12 teams the other teams being people like Mathematica in the University of Minnesota School of Health and things like that and little RPI little group of vets or undergraduates and we're we're submitting for the final on Friday which is why I'm a little bit watered right now and this this app is it lets users policy policymakers providers painters explore the regional disparity to spares disparities I'm sorry said a word disparities what disparities this crowd sources talk starting at the national level looking at both the the changes in mortality rate at a county level across the nation and then at individual state levels and exhausting those into high medium low risk groups based on their mortalities understanding trends most importantly understanding determinants social determinants which are gathered from a aggregate site called community it County Health Rankings which aggregates from 20-some sources and it allows us to understand the protective end and harmful social factors and visualize these and visualize these trends and explore and dive into it it also allows the users to get definitions of these different terms and it really is a it's designed completely on our as one would expect it's a it's it's done as a sort of a multi-page web app using our shiny framework it uses some JavaScript presentational front-ends such as full-page so it's got all the CSS good stuff in it but it's very cool and it's it's mostly done in real time so we're seeing these these correlations being calculated in real time and being presented and being interactive we've had to go through all the stress inducers like how does it work on a humongous display how does it work kind of tiny display how does it work on a and Chrome on doesn't work in Firefox all these all these different things and it's actually available right now a mortality minder 90 API that edu although it might change in 10 minutes when they do another merge I'm the merge master for the whole project that's why my hair is white I started off with dark hair so anyways it's been a real fun project it's been it started off this summer with the usual students that we bring into the data insight program math students biology students students who have been through that intro the data math program some of our past students from things like health analytics data lab I hope that Alex research lab but this fall we try something different we did an all-out call for students we put posters up all over saying you want a code you want to do cool stuff on change world come join us so we got a bunch of CS people know our experience and we decreed it's going to be done in our it's going to be done this way but it was a very good learning experience because number one and thanks to a couple people in this room who coached us at least three people in this room we had to put everything under tight source code management using github pull request a minute Mirjam every hour we had to accept in some other viewpoints how best to design these these interfaces may be using some JavaScript techniques here and there to be able to adjust things but also how we could apply with a lot of confidence sort of this pedagogy that we've been teaching for several years in the various data analytics courses so it's been very cool we're very excited to be to be actually completing something that we're very proud of we'll see how it goes if we win we get fifty thousand bucks for first or thirty thousand per second or whatever twenty thousand or something for third so well here's some time early in this in the after the beginning of the year what where we stood anyways it was a very fun project and and a humongous learning experience for the data insight program [Applause] alright we have few talks already about VSA so I'm gonna talk about this itself [Laughter] so PSA stands for dynamic spectrum access and behind this project we have the motivation of the US government so chunks of spectrum to the industry like AT&T Verizon and other wireless companies these too much use chants of the spectrum for their for their own mostly for military operations but for the users as well so what I did they created some set of policies to regulate chunks of spectrum based on some attributes like what device are using what frequency range of and what we're gonna do with this frequency from how so who does not miss exist so you know this project we are creating a a I system that is able to represent machine readable versions of those colors there are in natural language and but only that we also were also providing away or support supporting the role of the spectrum manager which is the person who's like allows or deny his mission by presenting him with explanation of why its mission God allowed or denied so they came understand and they can proceed with the proper adjustments to be able to proceed with their specific operational specific needs so right now we have a working system we have this project we have meetings every every six months what do we need to show what we have accomplished hurry from working system so I said I didn't have anything but of course that changes always ok so this is really ok it's like theta that's very strange ok so I I have this isn't what I'm working on but it's a question and I think it's a question that we might all want to think about but it's also maybe a question that we can think about as a research project and the question is what's so difficult about RDF right I think so so with so we all have we've all gone through the process of learning RDF to some degree I'm just easy sometimes it's hard sometimes it fits like really naturally in our brains like I took to it really quickly although the first time I heard about it I was like well why don't these URIs for everything but that's really offered but then once I realized why they made sense and I went with it so I'm wondering is there some sort of cognitive impedance to thinking in terms of RDF thinking in terms of ontology maybe maybe there's something there that we can either fix or address in terms of education but clearly there's something there and clearly we don't know what it is because everyone has a different opinion about what the problem is it's that sounds like a usability study that could occur that sounds like a entire cognitive science project of some sort that could maybe help us understand a bit more about what the heck is going on in people's brains and why doesn't line up with how certains brain works I don't know but yeah there's something there and I don't know if it's a full research project or if it's just people complaining but I think I think yeah and then he speaks five times fast okay so a couple things that I've done that are interesting over the last few months one is going to advise a bunch of funders about what kind of it's the United States should make with respect to artificial intelligence and the next generation and Wireless which was pretty cool and actually one of the things that we heard at this workshop was a competition run by DARPA do dynamic allocation and interestingly enough I think there were about 15 competitors and they did two big rounds of the competition and each time the winners were hybrid systems so there were machine learning only approaches there based only approaches and there were hybrid systems and each time the systems at one were hybrid approaches so we tried to give some advice to funders another thing that I did that was and there's gonna be one theme across all three another thing that I did was go to the Press Club in DC to talk about to a bunch of foundations and a bunch of high level of government people and high level researchers including the chief statistician for the United States yeah cool title well and actually this woman owns data gov so that's under her purview so the workshop was called rich context so what do we need to represent about the context of everything that we do about the data about the solutions what kind of language that the end was a bunch of social scientists it was behavioral economists it was statisticians it was funders and a theme so the first theme was kind of hybrid and be careful about the data that you use and work what was done to it this one is be thoughtful about the context in which everything happens in which the data is collected in which it's used and a third thing that I did just a couple weeks ago well as represent RPI and AI so I was the only person on the panel who was an AI had AI credentials and the only person representing RPI in a panel on artificial intelligence in the future of the Capitol region particularly with respect to cybersecurity clean energy and environment ending global climate change and actually one of the things that came out there also I made this point that it was that a diversity of opinions are incredibly important diverse teams are incredibly important data from all over the place is incredibly important context is incredibly important and no particular perspective on artificial intelligence alone is to solve everything so it's it's interesting as you move through the academic world and into positions where your voice won't be heard more about how you can help advise people who are in a position that really make differences in the world like setting policies and setting directions for research and I appreciate your support in this because none of us professors could do this without you and none of us could do this research without a full team that's a great way thank you