TWed Lightning Talks (Fall 2020)

Key Takeaways

[Music] welcome everybody to the uh fall 2020 uh virtual twed end of term lightning talk yay this is amazing um so the rules are are very very simple uh you have about about two minutes we're not going to make a big deal um if it's short if you go long we'll all start making weird noises at you and you just talk about whatever you want to talk about most people have selected something related to their research topics logistics we are recording this people actually like this recording because it's a great summary of the work that's going on across the tetherless world it's a real fun time i encourage everybody to have your video on it looks like i think most people do have the video on um and i would encourage you at the end of each person's time please give them the same kind of appreciation that you'd like them to give you and without further ado uh let's kick it off we'll go on the order that people signed up so that would be uh the alphabetically first amendment hit it thanks john um hello everyone i hope you're all staying warm today so i'd like to tell you a little bit about what i've been working on recently um i've been working with a network and graph embedding methods to come up with solutions to some of the main problems in the projects that i've worked on uh throughout the past few years specifically to do with natural systems uh things like datasets like mineral co-occurrence uh fossil co-occurrence as well as well as other data sets of um genomic data from samples uh so microbial communities and then along with the mineralogy or the geochemistry of the location and so data sets at different scales uh what i'm doing is trying to um using network embeddings come up with graph representations for these data sets because usually uh the information is more rich in the relationships between all the entities than in you know the properties or attributes of one of every particular entity and so it's really the interconnections between um between these entities in these data sets where the information is um so i'm trying to use uh really all the leading graph network embedding methods and to evaluate them on some common predictive analysis tasks to do with these data sets like mineral classification or like for example identifying communities or clusters within the fossil the entire fossil network of a lot of interacting fossil um families of different organisms um and so on with really with the objective of of taking the best performing which is at this point is the deep learning approach to create to come up with the presentations and enhance that by using some graph entropy measures to perform even better so that's it normally we'd have you go up to the white board and cross off your name [Music] all right andrew go ahead kick it off everyone again my name is anirudh for those who don't know me uh as you some of you might have seen on the topic that i signed up for it's the formalism to partition knowledge graphs into knowledge networks and as i was thinking this morning about how i'm going to talk about this i realized i can't really talk about the formalism without sharing my screen so instead i thought i'll talk about how i actually got to this topic how this actually emerged and then also mentioned that i'm presenting the same topic in two days time at agu and that poster will be published and so for those who are actually interested in the real formalism i'll put up the doi somewhere you know on twitter or on the tw site or something so this kind of goes back to january of this year a long time ago before you know we were actually allowed to go out of our houses and stuff but uh peter and peter and i were at the ecip uh winter conference and like we had presented this two-part talk at a open knowledge networks session and you know peter talked about what are open knowledge networks and how they're useful and i talked about how does one gain insights from them and i kind of used a lot of the examples of you know converting the data behind the knowledge graph into a network object and sort of apply similar things to what you guys have been hearing me talk about for a while now which is the network analysis network science etc so when i kind of talked about it a lot of people liked it etc etc but like a few people approached us uh and kind of asked is there a formalism for this can we generalize this is there like a way for us to take all of our knowledge graphs and convert it into these you know network objects uh because we really want to do this too and if there's a formalism if there's a thing that we can you can write up that we can cite and follow that great so uh so like that evening peter and i met and we kind of discussed some ideas and we agreed to go back to our hotel rooms and and just scribble some stuff and put some stuff together and then come back and discuss it so those scribblings those discussions that essentially were started off this formalism and we got like an initial idea for a generalized logical formalism mathematical formalism to cut to look at the data the sp the triple data one property at a time and then extract uh the data and format it into a network data object uh so that's kind of how we got here and created this formalism if you want more details please talk to me uh also the doi for the agu presentation will be out like after friday after i present so i'll post the doi somewhere and you can look at the actual poster we're also writing up a paper on this like a technical report so either ways there'll be a lot of updates coming up so that's about it thank you very cool thank you and i'm going to keep after you for that link all right brenda i want to i don't want to go first well hey everyone um first of all congratulations on being here and if you're able to extend that to actually participating in this round of tweak lighting talks congratulations so let's face it we've just worked through a semester controlled by a pandemic an unprecedented life event for all of us so just you know be gentle with yourself come off of video can you turn on your video uh believe it or not i actually don't connect my video to anything i'm still the non-performance so anyway um so i'm brenda thompson a third year multi-disciplinary science phd under dr fox these lightning talks are supposed to be summaries of our current research work and originally i was going to talk about my attempts this semester to master web development using drupal and drupal is an open source content management system that's been used as a backbone for science websites like deepcarbon.net however i worked as a ta this semester for compsci1 which made drupal far heavier to pick up than i anticipated so i would like to share a little bit of how my own research time and ta time came to overlap so initially i wondered how the semester would unfold for a class of undergrads who are experiencing their first semester of college in a pandemic without realizing that how the lessons that they needed to learn would also help me help myself so basically the semester and the research that i was able to do was a lot of ups and downs and frustrations for them the students as well as for me i spent more than my allotted time assisting the students assigned to me with things like navigating rpi where to go for additional assistance how to succeed in college and simply communicating trying to get connect with their own classmates and it took a lot of energy actually pretty much everything this semester took a lot of time as well as energy so anyway for my own work in creating the cyber infrastructure for some larger research projects here at rpi um using drupal actually kind of became a stumbling block to my own success simply because the demands of learning a new content management system takes time and focus and so just like the students i was taeing i was trying to learn a language a new format some rules all while needing to meet goals and deadlines and basically i spent far too many hours sitting and staring at a screen um frustrated by you know how much time zoom and webex and slack took but you know i managed um and all of this while i heard myself calmly encouraging my students to be gentle with themselves to protect their study and work time to set up templates to simplify their work and so on um and it took me a while to understand that i needed to wind up spending the equivalent amount of time implementing those behaviors for myself so ultimately my own heavy lift of learning and implementing drupal came down to just squeezing a few items out of my head and onto paper into templates and onto the websites all of this while hearing the students exclaim you know it feels like i didn't get anything accomplished but i sure was busy and that being said look for an announcement soon about some amazing new websites and thank you everyone finish what you can and be gentle with yourself hopefully you weren't uh using the tw portals as learning example code no thank you but hopefully we're using the dco portal as an example anyway fifi it's your turn go for it hello it's fiona i don't think she's on no she signed up next she must have thought i think it's at six because typically we used to have credits at six right yeah that's why i send out 14 emails that say it's at five um all right so jay you're next yep um yeah sure so i was going to talk about uh the study cohort extraction work that i've been doing um so this started with uh shruthi uh recruiting me as a urp to work on it uh because uh shruthi was was doing a great job on her study cohort ontology for um categorizing the different kinds of characteristics that you tend to see in clinical trial tables as they describe the patient populations and uh she recruited me and she gave me a bunch of pdfs and told me okay turn these into uh knowledge graphs turn them into uh study coordinatology knowledge graphs um that's what i've been chipping away at for the past year or so um and i've i've settled on a system which is a four step pipeline system where we take the pdf we extract out the kind of tabular information and then we use a couple of different heuristics to look at the uh specifics of the table so the first thing we do is we look at okay we have a table uh where is the indentation in the table and um since study cohort tables can have a whole lot of indentation and rows nested underneath other rows we kind of use that structure to say okay well um this row has you know three different sub rows and then they have their own sub rows and that kind of structure is actually a connected graph in of itself right or is is a kind of a graph and of itself of different layers of tables and so we extracted that out and using the kind of graph structure we were able to turn that into a knowledge graph just by identifying from the text well okay specific elements of a knowledge graph are kind of found in the table and how do they relate to each other how does the study cohort ontology say that they should relate to each other um and how can we put it all together um and in particular in the past little while i've been looking at okay if there's a biomedical term how do we turn that into a knowledge graph concept or an ontology concept and i've been using the ncbo annotator for that which if there's anyone here doing biomedical kinds of stuff um i can't express as much like anymore how helpful the ncbo annotator is for that kind of thing which allows you to just um give in a specific piece of text you can annotate it with all of the biomedical terms and whatnot that it corresponds to so um i don't know it's a really neat tool uh it's on bioportal um and i've been using that to basically given these pdfs you know generate this graph like structure and then going in and filling in all the details thanks j i almost blew the horn the timeout horn you quit just in time hi everyone um so basically uh i try to keep it uh short um um i spoke to you all the last time about basically uh uh um about the fact that about the fact about the fact that about the fact about the fact that about the fact that we started a new work stream uh uh particularly on particularly particularly particularly particularly particularly particularly focusing on particularly focusing on particularly focusing on explainability uh within the work stream uh within the work stream actually uh within the work stream actually we had a couple of i would say uh actually we had a couple of milestones uh the past year uh firstly we conducted uh um firstly we published uh firstly we firstly firstly we pub firstly we firstly we publish two book chapters um uh mainly on the foundations of and also um and all and all and also and all and also the directions for and also and also the directions and also and also that and also and also and also the directions for explainable knowledge enabled systems um post that we conducted uh post that um post that we conducted a user study with clinicians uh um um um um um wetting the explanation types we had found from the literature uh after that uh um after that after after that we decided to work on a formalism uh stealing the word from anirudh uh uh um a formalism for explanations uh um uh modeling specifically i would say um modeling specifically i would say um modeling specifically i would say modelling specifically i would say the system user as well as i would say the as well as i would say the as well as i would say the system user interface attributes related to explanations um we actually won the best paper award for that work at isbig this year uh and that was i uh and that was according to me the uh kind of like a i would i would term it as like a highlight for me this year um in this year which is so challenging uh that was nice for us um um currently now we're looking at um currently now we're looking at uh um um um currently now um currently now currently now currently now we're trying to work on a service uh um uh um a service to be able to map the recommendations uh to be able to map the recommendations from methods uh uh to be able to map the recommendations from to be able to map the recommendations to be able to map the recommendations from to be able to map the recommendations from sorry to be able to map the recommendations from ai methods um um into explanations into explanations into explanation into explanation into explanation into explanations which into explanations with inter explanations into explanations which can be addressed into explanations which into explanations with it into explanations in into explanations addressing user questions um that's about my work i guess and then personally i find myself uh being more empathetic this year uh um i've learned to be more calm and then because of that um i found um i found the pandemic a lot more challenging initially but then basically now uh it's become like i would say okay to be able to speak during zoom sessions as is for all of our zoom and webex sessions um and uh and i'm just hoping for all this to be done soon uh and to see you all in person um and with that i'll just say that stay safe happy holidays and uh stay healthy thank you all right thank you shruti all right matt hello everybody i'm going to talk about automating uh semantic data dictionary uh creation and kind of a progress we've made this year um so stds if you don't know them they're kind of a key component in a lot of our projects here they basically describe tabular data using ontological terms and can even be used to generate knowledge graphs doing that currently to generate them though you need to be not only skilled in ontological uh modeling but also have a working knowledge of the domain which often doesn't meet together and so we are developing several tools a ui to help build them and also a suggestion service algorithm to provide suggestions throughout one of the key components in this is the attribute suggestions so basically given a data dictionary description of a column so something like uh participant id we want to link that to the attribute that's closest in the ontologies that were given so we might want uh haskell original id or something like that uh and so originally we took a parsing tree methodology where we would take the descriptions and we would break them down to their components their sentence structure and try to use heuristics to identify which is that attribute and for the first few studies that worked pretty good but as we got more and more data in we found that there is a lot of variety in the way people describe columns and define uh terms within a data dictionary and so we've moved on uh to a machine learning model uh a transformer network where we basically use and the attention mechanisms built into this network to pay attention to keywords and uh find better mappings and overall this new methodology has proven better we i think we've improved by about 20 percent in our suggestions um but uh there's some drawbacks to models like this first in order to improve this further we need access to more data um and we're trying to address that now and first way we do is we talk to people together so i guess a quick advertisement if you've got stds lying around send it my way the other big one though is we're looking at using other data out there there is a competition called sun tab i think it's annual now but basically they have a similar task except without data dictionaries we're looking at artificially generating data dictionaries from ground truth to further train these models so future work we hope to improve performance there even further and then also extend this to other fields in semantic data dictionaries and hopefully enable non-experts to more easily generate uh knowledge graphs that the world can interact with and i think that's my lightning topic all right thank you matt thank you so much okay all right go for it okay hello everybody i hope everybody's doing good today um i'm going to talk about data rescue with you guys um not sure if the whole lab is familiar with the concept i think as peter's group this is probably a third or fourth project that i've been working on um you can think of data rescue as recovery or rescuing of data from say old journals or old pdf files or handwritten notes from scientists that are just shelled into filing boxes and that do not reach out to everybody for the research um especially given how much we are talking for open data or fair data this has like comment onto the forefront of projects that you are doing with different domains um one such effort that we are doing is with prebiotic chemistry uh it's it's with in collaboration with dr karen rogers or who leads the rare center up here at rpi and through data rescue you get to learn a lot of new things we've been doing this over the summer now um we were engaged in this activity where we have the entire research group whether it's professors phd students undergrads all working together uh reading papers like about 150 of them which we had curated and we were trying to extrapolate the data from these papers uh this figures and also um graphs and charts and tables but also data sets that were just in the literature and not in form of the formats that we as computer scientists or data scientists are used to working with so through this process we are trying to work with the domain scientists to understand what is important what is not at the same time collaborate with them and truly do a multidisciplinary project where both skill sets are non-negotiable you can't really do without the other and through this entire um project eventually we will get to research and analysis uh if you guys remember fong he did a great project uh like this even how had done uh with the i think it was in the deep carbon observatory and to some extent even i think a naruto's minerals um jammy project does tend to do that to some extent um so that's all about data rescue and that's my time i guess thank you thank you all right sola right yeah so what i'll be talking about today is kind of more on um the future research direction that i'm kind of hoping to follow so around this time last year for people who remember i kind of was talking about um work involving ingredient substitution and recipes and how i was looking into trying to substitute ingredients for reasons like say you want to make a recipe healthier and you want to reduce the carb so you replace your potatoes with cauliflower or you're eating them with some is allergic to something and so you need to replace some other ingredient to make it so that that allergen is no longer in the recipe well the higher level goal of those kind of substitutions really was focused on that aspect of health how can i make this meal healthier for me and if we kind of take a step back really then we also need to consider some aspects of context regarding what actually is healthier include the context of yourself so say any sort of diagnoses you have or particular allergy you have and whatnot or the context of who you're eating with or where you're eating say if you're eating by yourself this particular change to a recipe might make it healthy but if you're eating with a friend who has a different type of uh medical problem or diagnosis then that change might not really be healthy for them and so you might need to kind of reconsider what sort of changes you're making and so moving forward the the direction that i'm trying to take is kind of to consider context awareness in terms of a kind of personal context as of the particular domain the context of what goal you're trying to achieve so for the running example say the context of trying to improve your personal health and then providing some sort of recommendations based on that including recommendations for say individual changes like say substituting an ingredient or maybe also recommendations for like uh larger plans of things like say a meal plan that is healthy for your particular context for your particular dietary goals and uh to work towards this i've kind of been trying to use some work that i did in a class i'm taking the ontology and i was taking the ontology engineering class jeb of this term and so i kind of was trying to apply this to kind of the context of recommending courses to university students and using that sort of context and also for the funded research that i'm doing trying to do more of recommendations for people to make their diets healthier and such um so yep that about wraps it up for me thank you all right thank you so uh kara i think you're next all right hello everybody um i'm kara i am a limited phd who's been working with jim hendler lightning talk um and today i'm going to be briefly talking about what my thesis work was thank you um so my thesis was entitled exploring discrimination with artificial life and i decided to work towards my very long-term goal of trying to figure out some of the genetic and social factors that feed into negative human social discrimination obviously that's a really broad topic so i narrowed it down to a simple form of discrimination that we see both in humans and in many other members of the animal kingdom which is how similar you and your potential mate are on some sort of physical trait or genetic trait or in humans even some sort of social trait for example in america people tend to marry those who have a similar educational background to themselves and are less likely to marry somebody who has a very dissimilar educational level so that's something that we see very broadly both in humans and animals and in my artificial life simulation i replicated that by adding a genetic factor that makes me it's more or less preferable depending on how physically similar they are to that agent so i wanted to see if i could manipulate the evolution of this factor could i make it evolve towards preferring more similar or less similar mates with a genetic and an environmental factor the genetic factor i chose was genetic disorders and the environmental factor was the level of migration between two populations that were otherwise separate what i was expecting was that with my two levels of genetic disorders the higher level would lead to agents performing less similar mates because that would make their children more more likely to not have a genetic disorder which is very bad and life limiting and with my two levels of migration what i expected was that the higher level of migration would lead to agents referring more similar mates because then their mate would be more likely to be adopted to their specific environment what i found was i did find effects of both migration and genetic disorders it was only not quite what i expected so both the higher migration level and both the higher a level of genetic disorders led to agents preferring more similar mates so i'm still trying to figure out exactly what was happening in those populations but i did find an effect and i think this shows the usefulness of this approach i'm right very cool um could you post um somewhere maybe you could just email me you must have some papers that you've written your is your dissertation minted are you still refining it i'm sorry is your dissertation available or um or is are you still refining it you know to get signed up i don't know if it's available via rpi yet because it just got approved today but i can send it to you if you want just today exciting that's fantastic yeah i would i would love to you know that agent stuff is is um you know the agent-based simulation stuff is really fascinating to me i just would love to learn more all right uh severe go for it uh hi everyone i'm sabira rashid um i've been part of twc for like the last four and a half years or so i'm working with deborah mcginnis i'm recently a phd candidate with her um so today i actually wanted to take this opportunity to not talk about my research or any funded work i'm doing or anything like that but more so a idea i guess or like what really got me interested in ontologies in the first place um so um yeah so the title of this tweed talk i guess is visions of a semantic world so um in the world we live in there's a lot of there's people places and things essentially right and um but that list of things is very huge there's like organization there's uh processes events um languages um you name it and like um in terms of people people themselves have a lot of attributes they're humans and um maybe they have some identifier pointing to them um so so actually for a lot of the things i'm talking about um i want you guys to think about oh is there an ontology that exists to describe this thing that you've seen before or oh i've never thought of making an ontology for that specific thing so like um people for example have um contact information right so they have an email address or a phone number and there's like a name and that name itself can have a lot of different parts um they might have a social security number a driver's license um maybe there's some demographic associated with that person right so um there's a certain number of years old or they may have a biological sex or preferred sex or maybe they're married maybe um they're from a certain country what's their nationality or ethnicity or race maybe they follow a certain set of beliefs and so on and so forth like where do they live there they had there's a physical address associated with that location which has a house and inside that house there's a lot of things like um there's a lamp in that house or a television or a stove and you can go on and on in different directions um uh maybe they work at a certain store or an organization right so that organization itself has a address and like has all these products and um maybe they have a employee or maybe it's a researcher organization that sponsored um and then you can also think about like events um that the people are involved in so like this um person was born on a certain date um to these parents so um what else is in this person's house right so um maybe there's an instrument in the house so what's an instrument well uh i see the guitar over there or maybe they have a piano is it a keyboard a grand piano uh maybe they play some woodwind instruments like a saxophone or a oboe or clarinet um maybe they have a library in their house with a bookshelf filled with books what's inside those books um oh it's a math book or there's a physics book or it's a storybook and there's like all these things in the world which themselves everything's all connected to each other but it's so full of information and the the people in the world are like part of a society which has a culture associated with it or politics or religion so pretty much where i'm getting at is the world itself is very semantic and everything's connected together but you don't really see a underlying ontology connecting all these things but these individual things themselves could have ontologies on their own right um so i guess what i'm envisioning is this kind of like semantic world where everything within itself can be connected to itself but what i also want to be able to do with this underlying ontology or knowledge representation is to also visualize it kind of like the simulation uh care i was talking about where like okay now i'm in this virtual world where i'm driving a car down a highway on a certain geographical location and so on so forth that i can actually see maybe i can use it to predict trends in the stock market maybe i can make a movie with this kind of simulation maybe i can make a video game uh maybe i can use education you know so i don't know i'm just imagining visions of a semantic world all right all right thank you sir jason uh hello everyone i'm jason uh so i'm working with professor debra mcginnis as a phd student um so today i'm gonna talk about something related to common sense reasoning uh this is part of our machine common sense uh project uh so so common sense reasoning is actually some knowledge that people will not uh state it explicitly in everyday conversation or in the test so it's very uh because p people knows uh common sense but machines uh will not it will be very challenging for machines to understand common sense as well so uh the goal of the the whole project is to test to develop some approach or test has the ability of of the machine to see whether they can handle common sense so at the beginning of the projects a number of uh very challenging data sets was proposed to test this kind uh to test such capability of the machine but unfortunately uh as the uh with the uh uh development of a very large pre-trained launch model like birds robota the machine can achieve near human beings performance on those benchmark benchmark data sets and all those uh benchmark dataset is uh uh is multiple choice questions so um so uh after people uh trying uh is trying to propose some uh generated generation-based uh uh question questions like given a few words like describing uh a specific scenario and the machine will be asked to generate a fluent natural language sentence to describe those the scenario using those provided words so it's uh a also machina you uh asked to answer some um questions uh regarding uh numerical reasoning which is um some drawbacks of of existing pre-trained language model uh one uh one data set that uh we we are tackling recently is the uh is a question developed by a company called site so the goal of this data set is to test the machine about the ability to answer some entangled question such as the first true or false question sounds like a fish can ride a bicycle to uh and the second question is whether acquainted organisms don't have enough eggs to ride a bicycle so if the uh so people assume that if the machine really has common sense uh they could uh they yeah they should be able to answer these two questions correctly at the same time if they they actually can handle the background knowledge properly so uh so all these things is very challenging and interesting and i think it's necessary to be appropriately solved before we can achieve the general artificial intelligence so that's it thank you pretty big stuff thank you very much general artificial intelligence yeah mitchell go for it uh can you hear me yeah so um i guess i'll just start off by saying so i'm mitchell i'm a grad a co-terp master student with deborah and i've been working with with enrique on a basically extending the work of what the dsa policy builder did so as a little background dsa the dsa project brought to life a separate tool for building uh dynamic spectrum allocation policies uh that are semantic and semantics based so one of the benefits of semantics based policies is that you can uh reason over them and you can apply them uh apply logic so it doesn't so it can be read by a computer and operated on by a computer whereas a lot of policy language is often uh just a formal structured language but it's not machine readable so that's just like a brief overview and so the idea is for my research and this is going into my master's thesis is it's a a domain agnostic version of this so we have the the uh dynamic spectrum version of this well this well my version is going to be a domain agnostic version which can be used in many different scenarios and so uh currently i'm i've been working on it for the past uh i guess two semesters now uh because we're reaching the end of this one and so what i what i'm doing uh what i've got working is i've got the policy construction done where you can actually specify the attributes and all the all of the individual pieces of the policy using the language that was set out and based on exact mo and that in the uh dsa project using all of that as a basis for creating a domain agnostic version and the domain agnostic version uh is going to be like the idea is to make it capable of editing and creating policies as well as viewing them and just make it overall easier to use semantics-based policies for other applications so one of the applications that beyond dsa would be a an application like uh for instance like network access you can restrict network access you can like any permissions based structure works well but then there's also like other operations like uh like healthcare recommendations is one of them so if you anything you can have anything you can foresee where it's a sort of like you have a fun you have like a functional like i have this set of properties and only these requests can be fulfilled or these requests correspond to an output then you can make a policy that's really built on that so yeah all right so yeah yeah so yeah so mitchell i asked in the chat but um i'd like very much to get a demo um i actually had it running on the other screen yeah i don't know no i mean yeah no this is not the time to be something other than a lightning talk if you were to proceed on to the demo yeah i'd love i'd love to give a demo yeah yeah so i'd like to you know get um you and uh enrique together because i'm sure there's questions i'm going to be asking that enrique would answer you might not be able to answer but i i'd like to attack it and really get into it um when it's it's convenient for you guys all right so um so thank you for that mitchell um thank you and uh ishta hello uh can you hear me perfect please oh okay uh so hi uh for those of you who don't know me uh my name is ishida i'm one of debra's master students as well so this semester i have been working on generating explanations for food recommendation systems uh this work is kind of an extension of shruti's explanation explanation ontology work um where should these work derived and formalized certain types of explanations um using a clinical use case i am working on extending that into the food domain so that we can provide explanations for users um in for users of food recommendation systems so this semester specifically i have been working on a couple different types of explanations uh shruti's original list had nine i'm working on only three of them uh contextual contrapositive and contrastive um sorry i i think i misnamed one of those but anyways um that's fine yeah counterfactual um starting at the screen too long uh anyways uh yeah so i've been working on formalizing those in in a food domain specific way for example um if we look at uh counter factual uh contrastive explanations we would normally we would have two parameters uh for example two different types of food and we want we would want to find different facts and foils uh facts that would support a food or support a recommendation and foils that would uh suggest that we wouldn't want to go without so how do we formalize that um in and how do we formalize that logic uh in a food specific domain that's what i've been working on um we the way i went about doing that is an interesting mix of just uh of the ontology itself as well as putting a lot a lot of uh logic and inference into the ontology so my current um [Music] model abstracts out some of the parts of the of the um sorry extracts out uh some of the lower level food characteristics such uh abstracts out some of the lower level food characteristics such as um something that a user might like or where food would be available into these larger ecosystem and user characteristics um which i then infer on to get specific explanations uh next semester i want to extend this work into different explanation types specifically uh scientific and statistical um and possibly add some of this work into a knowledge graph yep that's it all right now we we have a number of other people and we still have some time we've got about nine minutes left in the hour uh is are there any staff type people who would like to jump in and say something we all have meetings all day long so our brains are perfect are burnt out well i'll jump in and do a little a little bit uh as some of you know i've had the pleasure i guess you'd call it of working with an army of undergraduate students since the spring developing a whole set of different data analytics applications to deal with coven some of them are some of them are visualization and data exploration some are related to a predictive modeling of the spread in in different localities and in specifically in different kinds of schools and some of them were tied into a reopening of campus in a couple different ways and monitoring and helping campus administration and actually students help with the de-densification of the campus and with this term we've we've come to a point where we've got like i think we've got like what five or six different applications that we've gone live with some of you have seen covet minder that was the thing we started the the year with which takes daily downloads of data from johns hopkins and a number of other sources and plots national trends state trends county trends and does uh determine analysis social determinant analysis as well there are two apps that that were originally conceived by malik uh malik magna nishbal and have been converted to web-based apps interactive apps including covered back to school which was used by the administration uh in helping to make decisions on reopening campus but it's been made a public app that any school that from primary schools to colleges can utilize for modeling different situations social distancing turning different dials and then there's safe safe campus and study safe which use aggregate wi-fi data to help understand uh what's going on on campus and and uh in different situations uh study uh study safe is really cool because it was conceived of this summer by a student and because we'll be it's in testing right now we're going through some user testing with it but that enables near real-time decision-making on where's a good place to study based on uh what tuesdays at midday are like et cetera et cetera and it's got all of these have maps in them and all of these have uh different kinds of analytics and predictive modeling so it's a uh it's been a real um a real long year i'm sick and freaking tired of covid um and covet data and plotting kovit kovit kovetkovit um and it's just going to continue but i wanted to just really praise the students that we've been working with in the idea data insight program this year who have done some real hard work put some long hours in to create some cool apps that are doing some good and that's my piece i have a question i was wondering how difficult is the cleaning data cleaning process from johns hoffman's website um so that's interesting so we have found based on our testing that the jhu data is pretty good we've only had a couple situations uh this year well at first i should say there's a couple different kinds of johns hopkins data that we get there's different streams you get time series data and some other data we have found at least two occasions where unannounced they changed uh the name they changed the column or they changed this or changed the format um which interrupts our our pipeline uh which we have to jump in and fix so we we have to there's a daily process of of making sure things work fine it's semi-automated um in terms of the completeness of the data that is um we do have to keep an eye on that uh our code has to be very tolerant of missingness uh in some cases it just doesn't matter since some of the stuff we're doing is like daily updates uh it kind of doesn't matter we do daily polls uh and sometimes it's self-correcting you know early in the pandemic new york would move around mortality the deaths would switch from one county to another county uh in the net that was not a problem because we would always do fresh data so but it's it is uh it is an issue and also there's there's a lot of data that's missing but they're really a lot of data about um demographics you know racial uh you know the data tied the race data tied to other socioeconomic stuff this is really important to understand the social determinants of health and uh and a lot of the united states particularly in the localities that need that data the most they're doing the poorest job logging that data that's not a johns hopkins issue that's uh kind of a general data collection issue with covid um but um anyways that's that's been one of the the the issues but anyways that's a good question we should be careful i'm not diving in on your lightning talks so anybody else want to uh say something yeah i'm gonna say something so i i've been working uh on a side project for a couple of months now so this geographical reasoning has been following me my whole life so i did a little bit of my masters i did a little bit of my phd and i did a lot uh on the sa as well so i i started to to create for now as very very minimalist uh front end for enabling some geographical functions in an easy way we have jury sparkle but not every triple store uh enables it uh jenna does a little bit of it and uh we have a new paper not wii but there is a new paper on this week this year is presenting a joey sparkle plus there is also a new uh work group working on a new iteration of geosparkle so uh i started to create this front end for enabling easy uh geo calculation uh over geographical data of course and uh so far it's very incipient but i'm uh slowly working on that because i think it'd be useful uh for us and uh for the community in the future so uh yeah that's just to say that i'm working on that sounds like a good side practical okay yeah all right anybody else uh deb would you like to finish up the hour here dab put down that glass of wine okay can you hear me now yes we can okay good yeah no wine only only coffee but the wine cellar actually works again it broke of course oh no oh it works again the cooler's installed and it's cooling again so the world is good yeah i just um i've been in a pi meeting a national institute of health pi meeting today which was like really fascinating and totally motivates our work uh we were all talking about combinability of data and all the challenges with that and how much precision you need and actually we had epidemiologists and toxicologists basically um saying they can't really do their work without us so super cool that's on the work side um but on the uh the bigger picture side i just wanted to express my appreciation and on behalf of you know all the faculty and senior staff at tetherless and idea we really appreciate what you're doing this has just been you know a year one of my facebook friends said she just celebrated her birthday and she said i want to do over you know i don't want 20 20. i don't want this to count so um you know i think we all need to be creative without ways to find joy and balance and be productive and be taking care of ourselves and so i just wanted to express my thanks because you know normally we would have a party and that's harder to do and um normally we would be uh you know having some uh food and i tried to find out whether there was any way i could have food and i got no no no no so um anyway uh there's not much we can do that we would typically do so i just want to say a heartfelt thanks and um you know we really appreciate you and none of us could do this alone we really are in this together both to get through covid together but also to get through our teamwork together so um thanks together apart yeah distance makes the heart grow fonder right all right let's everybody give themselves a rousing i'm not sure who was just talking amen was that you your network is really bad i want to know how severe and struthy and mitchell are doing these funky little things with their pictures yeah i want to go next to share there's this reactions button oh even the linux thing has it thank you i think that's new i haven't i didn't notice that appeared like today i think i'm surprised the linux thing has it they like almost never update really this is except i don't know does it show me doing mine yeah yeah you do we see you yeah i see a little quirk okay cool anyways with that all the presents are little pictures like that's very cool let's schedule another meeting just so we can play with this all right so thank you everybody um and uh take care of yourselves um this is the last sorry man you keep trying to say something i can't hear it um if you turn your i'm sorry okay go ahead i'm i'm sorry they're from not working i just wanted to tell everyone to that that spacex just completed their starship top test you should all go and watch it it is it's just amazing well the other thing that happened this hour was uh nasa announced the 18 artemis uh astronauts those are the astronauts that are going to on the artemis project to the moon and nine of them are women [Applause] so they've nasa has finally decided that they should send the smartest people in the world to the moon anyways with that have a great day all of you can say well it's about time it's about freaking time all right so there will be a recording of this uh posted in the usual channels when it finally when webex finally compiles it and all that kind of stuff so anyways take care everybody thank you very much