Quantum Conversations VIII: Sigmoid and Quantum Machine Learning on IBM Q

Quantum Conversations VIII: Sigmoid and Quantum Machine Learning on IBM Q

Recording: Quantum Conversations VIII: Sigmoid and Quantum Machine Learning on IBM Q

right we're good to start yes hi everyone uh welcome to quantum conversations number eight i'm mart if you are seeing me for the first time i started working with alexi on com conversations when i was at stanford uh now i graduated and managed the site together with stanford quantum computing association so josh is leading the partnerships at stanford now we'll be hosting researchers from sigmoid today sigmoid is a data solutions company they build and manage the largest state of platforms in the world they're also looking at quantum technologies uh it was actually their blog post on quantum support victor machines that made us reach out to them so today we will be talking about quantum machine learning and uh so usually we ask everyone to introduce themselves so we'll go ahead do that now we'll do a quick intro name affiliation um so affiliation i mean like school or company and what what what they're interested in quantum so maybe let's start with alexi and then go on to sigmoid researchers if that's fine awesome thank you mert uh yeah really excited for this partnership as you know i joined ibm i was a community partner right working outside ibm for years and finally kind of uh jumped in and obviously this meeting space is neutral and third-party so uh i didn't want to make it kind of ibm thing and so uh it's very appropriate right i needed to hand it off to community uh folks uh in terms of running it actually and so uh stanford quantum stepped up and merged graciously uh agreed to drive it so really excited for that because i think mergers like mercedes is the example of people we need to make more of during this process right like learn about it go to industry do it right teach new people and so i'm not already showed a bunch of learning resources and contributed a bunch of insights and i think this is like the main purpose of this how do we learn this and i just started the job right like within about a month and like i i feel like i know much less now like because there's so much stuff like the more i go into it like the less i know so so it never ends uh we have a bunch of you know a mix here so you know just you know gonna be helping from kind of community side uh right um and uh super excited you know that like there is a team now kind of leading this forward and we're supporting i will kind of uh represent ibm i will be supporting quantum conversations but you know it doesn't mean like we're open for everything like it's very important to understand i'm not the voice of ibm in the sense that i don't represent ibm officially i i represent the idea that community is important that's it so thank you guys looking forward perfect thanks alexis so uh ashish can you uh introduce yourself please sure uh thank you first of all thank you very much for uh calling us in uh and having as part of the quantum conversations uh i would say it was very opportune opportune and we got a message from alexey we were able to uh find some fan slash following on a blog that we wrote some time back it was long time back but we'll talk about that later in these conversations as well quickly talking about myself i am an iit kanpur passover it's an engineering technology institute in india i did my mathematics and scientific computing masters from there passed in the year 2007 since then have been working in the analytics and data science industry uh sometimes in the services industry some part in the ai and healthcare product companies as well currently i work as the head of data science in the vice president position at sigmoid analytics and similar analytics i think you already mentioned but we provide data solutions uh which encompass data engineering data sciences and uh as uh part of our job not the regular part of our day job we do some research and r d as well one part of that is some research on quantum computing uh just to give you some context and where it started uh i did my summer internship in uh university of mathematics india in france uh it was a three-month course which i did under uh professor dimitri patricides and therein we religiously used to follow julia camp uh who was from uh university of paris in france at uh emir again and she had done a huge amount of research in that at that point of time in 2005 on quantum random walks and that was my research topic as well uh which i did i think 15 years back i barely remember a large part of it but that is where the curiosity into quantum computing area began and uh we did put in some efforts at sigmar analytics to build at least some part of our knowledge in that particular area we were able to write some application based uh blogs which we'll talk about soon uh and those blogs are more oriented towards utilizing cascade library for practical applications including uh some analysis that we were able to perform on breast cancer data as well so that uh being said uh i would put a caution i would probably start with low expectations so we are not full-fledged researchers uh this is something that we do out of our interest and as much as time that we can gather uh we are no more than rookies or novices very much enthusiastic into counter quantum competing but no way experts so please feel free to shoot questions to us we might be able to answer few of them during these conversations or we'll certainly note them down for being able to address in future with me i have four of my colleagues as well uh we would probably ask them to introduce as well we'll probably start with giant giant would you so can i do that but i'll go ahead this is a free for all right then uh introductions from your site please yeah sure thanks ashish so uh i had the marketing at sigmoid and i specialized in b2b marketing and have been helping technology companies in uh building their brand visibility as well as creating demand and i'm really pleased to be uh participating in this conversation today thank you thanks chance and i think we have yeah thank you mark um so hi everyone uh here so like um i am a senior data scientist at sigmoid um currently so like and uh i have around uh an eight years of experience in this analytics and the data science domain um so that's pretty pretty much about the professional side of me but like uh coming back to the research part as ashish was mentioning so it's a great opportunity for us at sigmoid to um research on something very cool uh actually but this is something which i've never expected um which i would be doing um so it's like as he mentioned so his research some 15 years ago inspired or like directed us to start researching on on quantum computing as one of the research topics like we were also researching on a couple more topics but like this is something which we uh as a group like we were doing so myself uh amita manuel and um like we could uh actually like get into really cool things like um and practical implications of something which is very futuristic and everything so that was um great about uh this particular research activity at sigmoid yeah and um as she mentioned as well like so we we started like or we started a while ago maybe around eight months or ten months back or maybe before that as well so as as just an initiative and we could do a few some of the research we um we could uh like gather material across the internet and whatever is available explored a lot of things and like uh collated our research into a few few of the documents and blogs and as much mentioned so one of the blogs which we have written it is purely exploratory in nature which caught their eye and like we are also very grateful to talk about our experience and how we implemented in this kind of a forum very excited about it yeah so so let's we'll i'll go into the uh into the actual part of our experience and a few implementations which we did uh later so i'll hand over to uh mud to uh like ask everyone else to introduce them thank you all right thank you uh amit i'll make your own okay sorry guys hello everyone so myself amit and i'm a data scientist at sigmoid so i'm working with position vascular in many of the research projects and the business problems so i got a chance to work on the quantum computing case bhaskar in ashes has had already an experience in quantum computing so it was a very fascinating field and we got some interest in it and we didn't explore diving queue and biscuit library so we found some of the algorithms that were there already existing we explored them we went through some tutorials and all that and then we tried to implement it by ourselves and we were successful in that we tried to compare our classical we generally work on the classical computers and we generally learned uh classical machine learning algorithms so we want we wanted to explore quantum version of it and wanted to see how how the runtime would change so as a part of quantum research we went through a quantum support vector machine and i had implemented by myself the following algorithm and now i'm working in a big data architecture businesses problems and recommendations so that's all from my side thanks amit uh let's move on with rose um i have a uh local participant with me who is my roommate hey this is santi um can you say your affiliation and why you're interested in yeah for sure so i graduated stanford a while ago with a physics degree and i'm working on a little startup here with mark having to do with quantum computing and finance okay um alexander alexander would you like to introduce yourself hello okay um so we can come back to you um anirud anirud robot sorry if i'm butchering names by the way it's is also part of the sigmoid group and he is going to be one of the presenters for one of the blogs that uh we've discussed but please go ahead with your introductions yep so um i have done my bachelors and masters from iit kharagpur and there i had a not a couple of more than that subjects related to our electives related to the machine learning applications i have did a project there under one of our professors so that particularly interested me so just after graduation i joined sigmoid as a data scientist and after that i got this opportunity to work in this quantum computing research under the guidance of fashion bhaskar so that particularly interested me on the physics side also and based on the application it has numerous applications so we have tried implementing a couple of them based on the resources that we gathered so yep i'll be showing one of our applications other than qsvm but we are still working on it we have like uh done a version one kind of thing so i'll be showing you that in the follow-up session so yeah that's it from my side thanks mark yeah thanks um nayabi hello everyone i'm uh i have an msc number the systems engineering from the uk and uh as i am gonna uh apply for phd studies related to quantum technologies i am interested in to find out some research and some opportunity related to quantum technologies before studying phd or less that's it thank you um yeah hi everyone uh i'm bharathrav uh i'm working as data scientist in uh fintech startup in belgium uh yeah to honestly i'm here because uh yeah i'm brother of busker is one of the speakers here so apart from that i did my master's as well in computational chemistry in which i've gone through a course for quantum physics and quantum chemistry as well um yeah i'm just going through and and of course i work on machine learning applications so i'm just looking forward uh i just googled it that it's connected to entanglement and yeah some of the concepts i just want to you know link the interdisciplinary concepts between quantum physics and and machine learning i'm looking forward for that yeah thanks amazing thank you brian hey nice to see you again ryan lew hi mark nice to see you again as well yeah uh hi everyone as well as you all my name is prime i did my bachelor's in math and physics i graduated last year what a legendary year and now doing my master's in computer science i was really glad to hear that this call this talk will be about something that will have a more applied perspective on quantum computing so i'm very excited to listen to this hands-on talk and then gain some skills from that and so nice to see you all perfect thanks um go tom chander got got tom chandra okay um jamie gill nice to see you again jamie hey mark good to see you uh yeah just kind of like met brian on like the last call like two minutes ago oh we're just on the same zoom call too um yeah good to see everybody i am now at stanford at least for this quarter uh doing a few courses to sharpen my skills up uh so i keep learning quantum stuff so yeah um thanks for uh stanford to step up and alexi to keep on doing this because uh i've enjoyed all these quantum conversations over the last i think around what was seven eight or something like that right now they'll they've all been good so uh looking for another nice stay very cool very cool which course are you doing at stanford i'm in uh cs254 um competition field complexity i'm also um going through amir's course in quantum hardware is uh avia teaching 254 complexity uh lee yang is teaching it and then he's teaching um i might stick with that he's teaching 254b next semester and he's also doing a research seminar so i'm i've definitely been enjoying it it's definitely a good idea giant giant is also from sigmoid yeah hi i ought to introduce myself oh you did sorry i'm so sorry um john kustomski hey john you may be muted zooming etiquette you know um i have some experience in modeling i have a small company called tandoor data and i'm really interested in quantum where it's going to go and the spin-offs are going to come out of it if not directly but indirectly with all new technology my degree is in business but i got it because i was in i.t for the longest time in and out right now i'm just upgrading a lot on aws skills because as i get back into going for more data science i want to have the tools that i think are there that that jc just once you learn how to use the tools properly then you can just think freely and just play with it if you're willing to pay for it now ibm i haven't played with that much but in my experience with ibm up the street here in dallas essentially i've been i attended a blockchain and i was very quite impressed on how much their system worked properly because i i'm the kind of guy who likes chaos theory you know like what happens if i pull the plug in the middle of something and see what happens and that's that's counter light worked really well the recording worked and when i'm when you can trust systems that well in testing because that's in my early days back in electronics i was a test technician for like five and a half years so you that gets ingrained when you start to experiment what happens and are the um is the evidence that you're accumulating actually true and then you have to go back and test and make sure so you're constantly refining and the data that you've originally set up to follow your hypothesis great thank you um katarina peace to noah hi everyone um i am a physics phd here at stanford uh i've been working on nano materials and quantum materials for many years uh first electron transport and superconductivity and right now i'm doing more optics um yeah and i'll since i'm at stanford i decided to take a couple of machine learning courses just to broaden my perspective so yeah i just want to see what is out there perfect um oh hi everyone am i audible yes okay hello everyone so i'm i i graduated uh in 2020 and i'm working as a project associate at triple id chennai in the domains of deep learning and computer vision uh and i came across quantum computing around two to three years back and i started exploring about it but uh got to closely relate with it just six months or one year back and uh have have completed some courses and uh i'm here actually really to learn something and uh and and if possible uh if i can get an opportunity to work in a research project based something that sort so i really look forward for the session yeah thank you uh that's all hi uh my name is vasilyn and i'll drop in the chat some information for people who may want to get in touch with me i'm a mentor for the open quantum software foundation we just finished one batch of mentees and we are starting a new one what we focus usually is working with students and computer scientists interested on specific projects in quantum computing so if any one of you have a specific product that want to work i would suggest you can look up the program um and also i teach introductory requests to quantum computing uh for some of the affiliations of my institute for advanced physical studies in sofia or but i'm located here in california thank you essa nathan nathan shaw hi good to see you guys good to see you so i'm a good bit of an odd bird i'm retired always had a fascination with math and science i spent uh five decades in i.t sales marketing sales engineering and the like so uh i was really enjoying myself learning quantum math and then covet came along and given my age and uh stupidity unlike you guys who are smart i've smoked all my life so uh no one ever survived covered with emphysema so nine months ago i went to deep hibernation and obviously it's still working and it the quantum is i'm glad i did it because it saved my mind i've had something to focus on uh during all these times and uh i'm glad to report next week i get my second dose so uh we're all getting there by the way and uh i think by the summer time everyone will have it by the way just end it's it's in uh it's an endorsement to technology that they were able to pull off a vaccine in one year uh and actually one that works and uh for example with ace they've been trying for 10 years billions of dollars and there's no vaccine in sight so at least it's some light at the end of the tunnel there and good news for technology maybe with quantum around it wouldn't have taken a year would have taken two weeks yeah or two minutes all right good to be here thanks david um diana hi everyone i'm danny i'm a cs master student here at stanford i have some research interest in quantum simulation as well as quantum circuits and i'm curious to learn controversial learning right thanks uh ragavandra singh okay and i saw beatrice join in but she's still here she's also from stanford quantum and then josh is gonna be back soon so he can introduce himself when he's back but i can do a quick intro for him so josh is a sophomore at stanford uh and she he is he he started leading uh the partnerships at stanford quantum computing association after i graduated all right i think that's all did i miss anyone this is harder than it seems because zoom keeps moving around the list so if i missed anyone i'm sorry you can just introduce yourself okay so ashish do you wanna start with the uh with the talk thank you everyone sure okay thank you thank you everyone for uh joining in this is great to see such a good group of people there uh for our uh presentation i would request pascal to uh share screens we have been able to build a presentation do you have the presentation rights pascal yes uh we'll certainly try to keep this as much as conversational as possible but [Music] there's too much to cover from our side uh four speakers uh i would consider myself as half but uh we will use this as a visual aid for uh explaining whatever we talk through most of the talk will be done by bhaskar amit and anurudh they have been the ones who had worked on researching and building the blogs that we wrote i will quickly set the context for the conversation to follow uh pascal moving forward so from an agenda point of view what we would cover in today's session is a brief introduction to quantum computing uh some idea on the total amount of research that we have been able to do on quantum computing from sigmoid site so far uh then bhaskar uh amit and it will take us through what was their journey in terms of being able to explore tools technologies available in the market uh their journey in terms of trying to find out what hardware is available where and we can work to build or test its algorithms one of these stories which i would want to bring up is that so there's we wanted to initially uh work on d-wave as well but somehow we realized that the wave uh doesn't provide machines to run in the indian geography so uh obviously our first choice was ibm q and we were able to run most of our algorithms there but oscar will walk us to that journey as well uh amit uh would be speaking about the qspm implementation the blog which caught alex's eyes and that's why we are coming in here so that's going to be our uh core presentation we will try to spend most amount of time on that aspect and the la and last but not least and it will be talking about qgan now part one of that is something that we have already researched we have built or written a blog on that as well there is a follow-up blog which is coming up on this topic but it is held up because uh obviously we got caught up in some other things but that blog is coming up soon we would probably have it by the next month end quickly moving forward i would briefly talk about sigma uh basket on the next slide so what we have is a brief introduction to simoid we were founded in 2013. uh we're mostly a data engineering data science airmelo solutions company overall strength that we currently have is around 220 plus people working in data and analytics uh our growth has been phenomenal uh we have received a lot of awards as well primarily are on our work in the data science and data engineering domain sorry we are back then funded by sikoa and uh as moving on to the next slide basket on next slide what we have are the areas where we generally work in so these are some of the areas where we have been able to provide our services to our client set uh the client said as was mentioned in the previous slide is more than 25 plus clients in the fortune 500 group uh the business use cases that we have worked on include the standard but what i want to focus specifically on the rightmost column which is our r d initiatives so quantum computing is one of our r d initiatives we have another group dedicated to working on reinforcement learning and the application areas of the same our core focus generally is into marketing analytics and on separate notes we also work on gans cnns rnns and another area that we are deeply interested in is federated learning we feel that given the privacy laws which are picking up a pace in most of the geographies federated learning would become a core when we are trying to run models on individual devices without hampering or without infringing on privacy of users so even mobile devices this would become the way modeling or machine learning happens uh moving on quickly on to the next line basket and before i hand over uh to bhaskar to take us through the journey on quantum counting so far i want to talk about the quantum random box as well i think uh it is probably uh more than an hour long talk but uh quickly about the glimpse of the work that i was able to do during my internship so it was uh it was supposed to be twofold we wanted to work on building a string search algorithm in a quantum space and we also wanted to build a graph search algorithm uh which can work again in the quantum space uh we were able to implement in a theoretical sense uh or prove the capability of a quantum random walk to be able to reduce the hitting time for a certain length of spring a string in an infinite uh length of string so the core problem was that uh let's say we had a string uh we had named it as abracadabra and we are trying to find that string in an infinite length of alphabets uh so it's a search algorithm which works in classical sense it would take a polynom polynomial amount of time to converge uh so the waiting time to find that string in an infinite uh length of alphabets would be polynomial in terms of the length of the string that we are trying to find with the quantum uh random walk applications and uh in cmc kind of simulations we were able to record it in logarithmic time as well so that was theoretical we are yet to implement it uh in real world but uh overall uh there is a research with some which was something uh that i wanted uh to do more of but probably that point of time i was more and more uh fascinated with france than quantum competing but uh leaving that apart i think uh if we get enough time if we get enough preparation we would love to present the quantum random walk in our paper at that point of time uh as well in a separate session our core focus is going to be the blogs uh that the team has written uh we will probably land on the part in 10 to 15 years from now please bear with us and i would ask bhaskar to briefly explain about our journey on quantum competing and the tools and technologies that we've been able to explore so far i would pause for a minute for any questions so i have a quick question um so you said you implemented the random walk uh work on theoretical um grounds so what did you mean by that and if you have a paper i would love to see the paper so can you share the link on on the chat yeah certainly it's not a published paper if it was published i would do that but the uh draft version of it i have i would be more than happy to share with this group after this call but that's okay perfect thanks certainly we'll do that and i don't see any questions but please feel free to like shout out questions if it's very pressing but we also have a um questions uh session in the end after everyone speaks okay thank you ashish okay you guys can hear me right yeah okay so like starting from where as you stepped over so basically i would take the part of um walking uh you guys with with a summary or like a generic overview of what we could uh accomplish in our research in our short tenure of our research so like uh the the core idea of where we started was to uh was we wanted to explore um quantum computing its applications and the our our ultimate motive was always to uh look at the applications in the machine learning domain because that is what we generally deal with and quantum machine learning is something which is a very very um like something which is uh highly look forward to uh by a lot of um like experts in the in the field uh the domain as well so we wanted to um get into the um like applications of machine learning and also the practical implications for how how ready the current quantum world is so that was also one of our important um you can say like points to look into so that is how we started our research wherein we uh started exploring the concepts of quantum computing uh so as like i think in the in previous quantum uh conversations uh over here so i went through few of them as well so i think there were a lot of talks wherein few of the concepts were touched upon and there were also also a few of the talks wherein how should uh or how can someone um learn quantum so i think one of the sessions was that was great actually so maybe like if we if we had something like that at our disposal we could have done a lot better job in um getting this uh knowledge at a much faster pace and all uh also getting it at a much deeper level as well but like we used the resources available across the internet uh for uh to understand the concepts of quantum computing and um also like and during this uh during this time when we started our research like even in a where we were in a nation stage of our research there was a kind of a bombardment of this particular news wherein google came up with a quantum supremacy experiment i think a lot of us remember about that so that is something which triggered the entire um like internet uh in a way it shook the internet like and there was a talk of war between the giants about how can someone reach quantum supremacy and like we were uh like we were not uh sure of what corner supremacy was and we wanted to explore uh more about it so that is where uh we want we actually went all in into uh this uh wherein like we exploited exploring multiple tools we then we got to know that there were a lot of tools at uh at our disposal wherein we can actually run uh quantum algorithms on a quantum machine or a simulator which which is like a supercomputer which is designed to do uh design to work like a quantum computer so that is something which uh when we heard about this and then we also went into multiple research papers uh wherein a lot of like a lot of research has been done on this field that is uh something which we uh came across wherein uh we saw uh like few papers on quantum machine learning um like papers starting uh in 2010 2014 wherein uh people started like how can someone uh convert or a regular machine learning uh algorithm into a quantum um like a its quantum uh self at times so and what can it be uh can uh its applications be so all the things so those are our initial research and documentation which we did so then like as soon as we went into that so we um we wanted to like get into uh the exploration part of uh the tools and libraries available so i know like in in the previous quantum conversations as well and um like and uh currently there are any number of tools available a lot of companies uh a lot of organizations uh like um and also like one of our uh like attendees also mentioned there's an open source quantum research as well so a lot of tools available for uh for people to who are interested or enthusiastic about the quantum space uh to explore and dwell into the space but like we did some uh in-depth analysis into the available tools and libraries which we could um like gather at that time uh anyway so and we compared the pros and cons of each and every uh tool and a library based on ease of usage availability and coverage of libraries and algorithms so that was one of our important parameters we we used actually so as ashish mentioned earlier so the first thing which we are the first uh company or like organism which we reached out towards d-wave because that was kind of like they were trying to build quantum machines and they were ready to offer it to uh like anyone for free so we reached out to them but unfortunately we were not um given access in the asia pacific region so basically then um our alternatives were ibmq and google search and um like amazon's like one of the uh like you can say a tool something which we could not explore uh in depth but these were the two alternatives which we had in front of us and then we explored them and that is how we chose ibm 2 which which we will get into uh some time later then we also went into the applications of uh qml algorithm so qml is like a quantum machining algorithm so we wanted to see like uh in the libraries of a few of these libraries there were um like like uh pre-coded you can say libraries which had a quantum machine algorithm accorded so like a qscm or a q-gan or a variational quantum uh like classifier or something like that wherein we generally work on classification and clustering problems uh like for clients for business to solve business problems so we're very intrigued about how can these um algorithms uh work the quantum space and uh we also wanted to select specific problem statements like a classification problem we uh we also explored uh the breast cancer classification problem uh or like because there was a data set available in the co in the discrete library actually and the quantum svm algorithm which is which works as both classifier as well as a regression uh technique so that was also available so it was and it was related to breast cancer which is um which is much closer to my experience in my previous company wherein i we i was working for a healthcare ai firm where we used to predict patients um like you can say probability of uh getting uh like diagnosed for a disease something like that so we used to um do uh all sorts of um like pro or solve all such problems like this so this kind of increased us so like that is where we started um like uh using or like you can be exploring these um like uh tools as well as the algorithms inside these tools and libraries so once we went into these we then explored different functions syntaxes variables and um there this the libraries were very fascinating in a way so they have multiple branches and multiple you can say uh like divisions for each and every specific application like example machine learning was one of the uh sub sub sub sections of this particular discrete library which we went into there were many other applications which we were not aware of on the first place actually so it was very fascinating to see there are so many applications where content computing can be used that was one of the fascinating parts but like we we try to focus on the machine learning part actually like because we were that is what we focus we were interested in so like we then we tried um like taking some existing code and implementing it on top of uh a data set and uh running it on an actual machine actual actual quantum machine as well as a simulator so this was our um like our work actually in a way and we edited tested the codes as well there were a few syntactical changes from a lot of resources we collected so we actually uh tested that edited it and we also tried uh submitting few pull requests in terms of the syntaxes to the like uh to the branches actually but anyway like the now if you see um like as a interesting thing is that the entire ibm queue and the kisket library has gone into a complete overhaul that we ourselves cannot identify where is the thing which we actually worked on so it's like and they are doing a great uh i mean i can say like it's an amazing work actually their vision is great um and the way they have changed it it is it is much more easier now so for anyone who wants to start um using or like implementing any quantum related algorithms or like wants to learn quantum we would suggest like that is a very good platform ibmq as well as kisket so that's what like we'll come to those like smaller details about those uh very soon and the last part of our research here was like publicizing our blogs like whatever we worked on so we just wanted to put them in blog so that at least like a common folks and our friends as well like could uh go through them because um this is something which is uh everyone's like interest always at the back of their minds they might not know like uh they are interested in it but as soon as they see some content like this they'll be very much interested so that's what we wanted uh to um to trigger actually so we wanted to put everything what we did in a blog so though the blogs are not very much in detail but let me try to maintain some amount of like depth in it so which which we will get into later part of our presentation so we will see um so that is what uh as a summary about coordinate computing so now like just quickly talking about the landscape which we uh landscape of the quantum computing research so this is not the landscape across the world or like in this entire space this is the landscape which we could see or the horizon beams we could like look at so we just split it into three three things again research topics the tools explored and the algorithms implemented the research topics were mostly about the quantum supremacy experiment getting to know what actually google did and what what uh maybe like ibm was um accusing that like this is not true and again there were a lot of um like it was it was a tug of war actually it was interesting at the same time so basically we wanted to get into uh the details about what actually quantum surprise x-ray because we also published the basic understanding of our uh of us in a blog as well then we went into coding on quantum computers understanding the circuits and gates and basic principles of quantum computing by the superposition entanglement and everything and we also went into some research um our like basic tutorials about uh standard algorithms like the schwarz algorithm or the bernstein uh vazirani algorithm so which was uh like which was readily available in the in the cascade um tutorial uh youtube series actually so it was pretty intuitive and then we also wanted to implement a few of the kernel-based algorithms because of its ease of implementation in a way like um that is what we start so svm with a support vector machine which we will get into later and knnk means pca so all these are basically kernel based algorithms which can which have a higher practical implication of a quantum uh side of it in a way if uh so and also quantum neural networks and recommendations are ultimate goals so that is our um so our little farfetch'd goal is to publish a paper about uh quantum random quantum recommendation engines um that is one of our fastest goal in a way which we will hopefully will achieve soon so that that is more about the research topics we covered and the tools which we explored um like uh where these the ibm queue uh and cascade library together and we also went into d wave we exploded but as maybe we um faced uh some challenges with it so google's work and uh tensorflow quantum was also something which uh which we explored but like they they were not as uh user ready then when we were exploring like we we found ibm q and q to be very much uh user friendly at that time and right now as well but like the other platforms also have amped up their game like every platform right now is very much friendly so i think um it's every every platform is a go to platform right now but back in back then when we started we felt um uh bmq and cascade to be the most user friendly and also having the algorithms which we implemented like on the right side like which which you see the quantum svm for classification the quantum gang so both of these algorithms were readily available so we thought we will try uh just like a scikit-learn um library they have these algorithms in mentioned in there or like written in their library and which we can call upon and implement it on any of the game so like how to use and how to implement it is something which we have to learn but uh at least like it did the part of coding it out for us right from scratch so that is something which is great and there are also in our to-do list we also have quantum pca quantum k means quantum regression as well in a way whereas these are are not actually currently available in this library so this is something which you wanted to do or which you wanted to implement from scratch wherein you want to come up with a circuit um and to like calculate a quantum pc and contact a means uh clustering and and implement it on an actual data and see how it works so that is something which we are planning to do so that's our landscape yeah and um again so this is basically i'm not kind of uh both watching for a specific group here like our ibm as alex mentioned he is like he's a part of ibm but like we are nowhere related to ibm or cascade but we found this so user-friendly that like we wanted to um like uh watch for it like at least in this presentation but just a brief introduction about ibm q and cascade so ibm queue like is a ibm quantum that's what it stands for item quantum experience so for you for someone anyone like to just uh delve into the world of quantum so this is one of the best starting points you can see so they have the only thing you need is a ibmq account this can you can create it through your google account itself so you you the best part about this is that you get to work on actual quantum systems so there are nine quantum systems available for uh anyone for free actually and there are there are 11 more which are paid i guess and um earlier there were very smaller machines now they have a two qubit three qubit five qubit and 16 qubit machines as well um which are placed across the world and each systems uh differ in qubit topology a quantum volume or a number of qubits actually so basically you can see which gates are specifically um like configured in a particular machine and you can choose that machine for any specific algorithm which you want to run so that at that level like uh the uh the you can say the details have been mentioned there and um like coming you have to kiss kit is it's an open uh source framework um it's similar to any any algorithm or you can say sorry uh any algorithm a consisting library um so a co quiz kit is all about like circuits uh the quantum algorithms like how does how do you access the hardware how do you trigger a simulator how you trigger a quantum machine uh how all those and like how do you uh like get into a result or you can say measure something so all all possible like um quantum applications are coded in gift kit so this is the most important or like the um it's a heart of the this entire thing also like our main reasons for choosing this is that like we can readily have just called psvm and you can uh like like as well here which we the the classical part of which we are kind of aware of so we wanted like why not try this that was our main motto that's how it is okay so yeah um then coming back to like so this particular slide is just um a preparatory slide for a for the slides coming ahead actually so i just like put across um like a few pointers about the quantum gates and quantum circuits so just before that just wanted to touch base on uh like very very uh briefly on the quantum properties actually so this is where we started off with we understood what superposition was so wherein like it was kind of difficult for us to understand what superposition was um because like um it is very difficult to visualize something uh in the quantum realm so as we see like um something we everyone's um pretty uh like no or everyone must have watched the ant-man movies wherein the entire movie like i the only word i could uh listen to was quantum so i was kind of fascinated by then but like we could not understand like what exactly um it related to so then like understanding these properties and going um into depth of it like like the superposition wherein a qubit or an electron you can say can be in the states of or can have an upper spin or upward spin or lower downward speed you can say so this understanding this concept was difficult for us initially but like later on uh this is something which we could um like understand in terms of the practical uh machines which were being created like few machines are based out of uh based on electron based machines you are uh you can say the charge is something uh like maybe superconducted uh materials would be used so basically the charge is a positive or negative so so that's what we understood like how uh something a cubit can be in in superposition and um we also like understood like what entanglement was like when when you have one qubit so basically qubit is again like a building block of uh quantum um well you can say so so if there are two qubits or two qubits can be connected to each other and both the qubits can be only be measured without uh without one another so that's pretty much in in a single sentence if i have to summarize entitlement and this was also very um like important wherein like uh this can be or two qubits can be put into entanglement using few quantum gates so that was uh one more information which we could get so that is again a basic thing which uh someone would go through uh while going through quantum computing so quantum gates and quantum circuits these are the basic building blocks of any anything like any algorithm which we have to build you know be it upon the machine learning algorithm be it uh optimization algorithm be it uh like you can say um anything any kind of an algorithm um or a mathematical solver or something these are the ones which someone has to focus on so there are like one qubit gates multiple qubit gates so a gate is as simple as a classical gate so wherein it's an and gate an or gate that is a classical gate when it comes to a quantum gate there are multiple gates which you can um like which which there exist you can also create your custom gates in a way uh using ibm um there is a there is something called as quantum circuit uh builder in in ibm cube where you can position your gates uh on the uh on it and like see how how would a qubit change its its uh state because a cubit can be as as we talked about in a superposition so it can have a probability of both zero and one so it can be in both the states so you can change this space you uh using these gates so so all the quantum gates are reversible and they are usually represented by unitary matrices so for someone who is interested in understanding the calculation behind it so matrix representation is very easy for someone to understand it so this these gates can be represented in the in unitary mattresses and multiple qubit gates uh exist as well like uh like the swap gate or the c naught gate or the c z gate wherein um when you would uh like work on like at least two qubits to to get your output so basically if you if you are in uh in any kind of a gate there are uh there there are like two uh if there are two qubits and the output also would be 2 cubic so the number of input qubits is would be the same as the output qubits of any gate and you can say like quantum circuit is again like is the main thing like where in which would transform or manipulate the qubits so just like um like a sequence of ones and zeros is manipulated by a classical computer using gates um the quantum circuit manipulates the qubits this is what uh would um change the the quantum state from an initial state to a desired probability state and also this is what uh is required to change or like convert a classical data into a quantum data engine so that that is like pretty much about the quantum gate uh and circuit and uh like yeah one of the last things just like just to touch upon this i know this slide has a lot but like uh effectively it it is mostly related to the later parts uh part of the presentation so quantum machine learning this is just a very very easy way of putting it for uh it works this is something which will be gathered from a lot of research papers actually so what we see is like a quantum machine learning algorithm is like is a quantum algorithm which is as simple as like which can solve the typical machine learning problem but can also use the efficiency of a quantum computing um like uh you can say for the computing in a way what what do you mean by that as example if you have if someone wants to uh like work on a support vector machine so the the quantum approach uh would help us calculate the distances basically because uh all these like k-nearest neighbors support vector machines and k-means clustering have some other way um like an integral part of it would be calculating distances between points in any any n-dimensional space so this calculation is something which can be uh like sped up uh like by some some super something enfold through this quantum approach so that is um why this quantum machine learning is kind of uh is one of the things for the future so similarly for neural networks and decision trees we see the first explorations of quantum models like wherein that is what we uh we have mentioned here lies a quantum approach but like a lot of um like you can say like leap in this particular area um like and on to the right this is one image which which we are very of uh fond of in a way so this was something which we found on one of the like github repositories of some person like this was across across the internet as well so this particular image gives us um or like beautifully explained to us like the um like the correlation between or like the intersection between the machine learning realm and the quantum information processing realm or the quantum computing so the circle or the you can say vowel inside it is the quantum machine learning wherein quantum annealing quantum topological algorithms all all the algorithms which i just mentioned the quantum pcs vms and everything and um controlled methodology for methodology related algorithms all of these would be a an intersection between the machine learning side as well as the quantum uh information processing site so this is pretty much the uh like the quantum machine learning related uh like you can say theory you can say so now uh what we um could do was we picked up one particular algorithm like it can it could have been anything but we picked up svm because that was really available in the library and then we went into uh implementing it and looking at the practical side of it though we could not get an actual obviously we are still with the uh like noise we are we are far away from the fault tolerant quantum machines kind of uh stage so but still like um we could we wanted to understand the practical implications of any uh quantum version of a machine algorithm so that is where we picked up a qscm and a q-band so i think like i will like uh amit to take or like take you guys through the actual implementation we did on the quantum spm and quantum quan uh actually and uh like uh yeah android will be taking you through the quantum question implementation so um quickly over to you amit uh and before this like if you have any questions we can take it now or we can take it at the end of the session as well good question he's good oh sorry sorry interrupt hey hey hi it's uh kevin roney um i'm just there's just so much uh interesting stuff happening here so many great interesting frontiers and the engineering uh cutting edge on these quantum systems it's it's evolving rapidly it's just it's unclear to people who are uh not full-on practitioners which specific algorithm especially within the machine learning domain will make a supremacy style a breakthrough with with broad commercial applications because there's many different options here it's it's unclear who is in the lead to do really the big commercial breakout what do you think well that is the big question right i don't want to but yeah i would love to hear about this as well yeah yeah that's the big question yeah so again like this is something like um who is in the lead in terms of algorithm that is what you want to understand or in terms of name but no no rather in terms of the specific domain of application down to a specific algorithm i mean it could be svm it could be pca um i don't know but i'm just it's hard to uh hard to see the landscape with clarity now yeah okay so just as hard it is for us to understand it's as hard it is to explain in a way i would say can help me answer this but i would say that all of these uh like you can see all of the algorithms currently are not practically usable in a way as of now as as per my knowledge so we could on we could only just see or implement a couple of them there are a lot more so uh svm is one of them in a way like that is one of the classic classification algorithms you know it can also be used for regression uh actually but the thing is practically if someone has to use it to achieve a supremacy or like a or a time advantage over a classical machine i i it highly depends on the hardware rather than the software so hardware is not that ready according to me but someone can correct me if i'm wrong that um like there is so much to do about the about the measurement part of it the fault tolerancy part of a quantum machine and also if you see um we currently are dealing with a very small sized qubit machine like maybe 10 15 16 or google's corner supremacy was achieved on a 76 or maybe 72 qubits of a machine but the thing is for something to be implemented um for a business application in a way uh or an algorithm to be implemented we need much more qubits than that because the amount of data we have is also huge so like the google's quantum supremacy was just to generate a random sequence and replicate the sequence that that is not an actual practical implication um because a quantum svm can solve um like some a patient's prediction of maybe cancer or something on a classical machine and that would know on a huge data set but that is not um that that easy in on the quantum side so basically none of the like algorithms are in lead you can say in according to me but like if if there is something else if she wants to add or someone else wants that they can sure i think uh yeah kevin i think that's a great question i think everybody is trying to find a solution to that or an answer to that what's going to prove that quantum computing is worth it uh from our experiments as well uh what we wanted to test was that uh does quantum computing based analysis brings in real uh speed improvements efficiency improvements we were not able to observe it so far uh i'm a methanol to cover that in our blog analysis as well primarily because uh the major part of the computation went into converting the classical data into quantum data as well but and being able to run quantum algorithms on top top of that there was something which was a challenge now addressing your question i think the first thing that i see the area which would be impacted uh the most i think it's going to be the encryption decryption and the password and policy protection uh privacy protection that's the area which is going to be impacted fastest uh or earliest because of quantum computing and again it's a bad way to put it but i think it will start because uh hacking would become very easy when quantum algorithms start functioning at their potential and then obviously uh the the other part of it which is being able to make more secure passwords make more secure apps that area would pick up very fast uh i don't know can't say certainly but from my point of view i think neural networks uh are the areas which are going to expand so the complexity that we see currently in neural networks which makes it uh unobservable or unidentifiable so far would expand further and that is where the quantum competing algorithms and calculations the complexity the computational uh is would help best that's our hunch we don't know we are also looking forward to exploring that area i wish you luck a fascinating question thank you yeah let's please come back to this story this is a great discussion question and it would be great to hear everyone's thoughts on this thanks kevin i'll continue on the quantum svm including thanks vasquez for giving the oven landscape on quantum computing basics and quantum machine learning and thanks again for joining and giving us a chance to speak on the experiments that we did so quickly checking with the group as well are we going slow are we going fast are we uh covering too much basic stuff please let us know uh we will ramp up on that we're doing good um i know fine it's it's good if we start this cut the discussion around five maybe five ten but um if that's fine how many minutes do we have to cover our presentation so maybe 20 minutes to half an hour is that enough okay i'll then quickly move uh walk it through our presentation and let me to give this so what quantum support vector machine is it is a quantum enhanced version of the support vector machine and what is support vector machine is it a supervised learning algorithm that we generally use to classify the data and sometimes we solve the regression problem with it also right but mainly we'll keep our talk around the classification part only so let me brief you about how the support factor machine works and there is no much difference between the working principles of the support director machine and quantum support vector machine the only difference that we face is the calculation part and the computational expense expense that we are reducing using the quantum phenomena so support vector machine works on the principle that it classifies the data in the suppose the data is in the two dimensions so it tries to find a line that will separate the data and there can be multiple terms of lines that will separate the data so how do we choose between these lines is uh the question so what the model learns you know and tries to learn is it creates two lines uh we can in limit terms call it as a tunnel uh on the either side of the tunnel the data point will reside and then no data in between the tunnel so the model tries to maximize this margin and as is the model iterates through the data points the parameters of the hyper planes are determined so after getting the two two uh hyperplanes or a say in two-dimensional line we get an average line uh between these two which would certainly be the best line that would classify our data pointed to two classes so when we speak it in two dimension it is it looks uh so simple right but what if we took uh real if we see a real word problem where the data set is too much complex that it is hard to separate it uh using a single line so in that case we can't use a linear sem of course so we go with the kernel scm and use that uh kernel tricks that is readily available for the kernel algorithms is vascular has already mentioned that we use kernel algorithms wherein we only want to calculate the distances between the data points right so in case of scm also the objective function that you are going to see here is calculating the inner product between the data points in the hyperplane different data points and uh learning uh to try try to minimize this objective function so as to get the parameters of the hyperplane so what if we uh the data is not linearly separable in the given dimensional cell space so what we need to do is map it to the higher dimensional space and then do the inner product in that space right so it can be so much computationally expensive if the dimension of the transform space is too high so as we can see uh that phi is a function let us suppose that file is a function to map the data point from the input spec two dimensions to say three dimension then the cost will be not not that much but if we say about the dimension would be 10-fold or 50-fold then the computational cost will be so much just because first we need to calculate the data points that are in the original space and map it to the higher dimensional space and then calculate the inner product in that space so to reduce this computational expense expense of the algorithm what we introduce is kernel trick kernel algorithms con what kernel does is it calculates the inner product explicitly without visiting that space and result the same scalar value as the inner product would have we would have get in the higher dimensional space so this is a kernel tick that [Music] that is there which will reduce our uh tension or we should reduce our task computational expense so much so this is the basic uh of uh support vector machine now we will come to the quantum [Music] version of it so coming to the quantum version of it uh there are three basic steps that is in the quantum application or in quantum algorithm that is processed so first we need uh for performing a quantum say quantum operations we need our data in the quantum format either we generate it from the quantum machine or we can get it from the classical data and we can transform it by applying some circuit and transform it to quantum quantum data then we do some processing and the processing part in can we can involve some doing some pca doing some operation logic gates operations so as to uh do the task of the algorithm that we want to run so right in case of the support vector machine we want to classify the data so that processing part will be the step two of our high level overview and then we take the measurement in the higher dimensional space uh so as to get the classical output so quantum support vector machine does three tasks first it converts the classical data into quantum data do its some processing part and then we measure the output so that we can classify the new data which is unlabeled so coming to the uh practical implementation so what do we need how do we proceed with the uh you know implementing our algorithm or how do we write a code in library so the basic implementation involves uh creating a ibm q account what we did created in an ibm account use key skit library used its inbuilt functions and feature map that is feature map was used just to convert the classical data to the quantum data then we uh we use the breast cancer data set which was around 569 records and having 32 parameters so we can't run all the 32 features on the quantum machines but for because in order to run that we we need to have 32 pivots and that is hard to get at this point of time so what we did was we did some pca reduced the 32 features to two two features then we selected two qubit machines then we uh defined a feature map with second order expansion as we we had only two qubits we made the entanglement then we uh created a quantum instance although we created we called a quantum estimator which will be uh training then in order to train the quantum instance we need some backend execution execution settings that is uh there in the uh ibm queue skit library so we selected backends as a i1q simulator and also at the same time we ran algorithm on the ibm queue machine so in order to run the algorithm we have created an estimator we have bring bring out a quantum instance using these two we can easily input a train and test data set into the estimator and we can run an algorithm so algorithm will run learn to classify the data by learning the weights of the data points and at the same time it will give the you know importance to the some of the data points that that would clearly affect the equation of the hyper plane after that once the data set training has been done we can easily make the prediction using the quantum instance in qsum.predict function that is there in the quick discrete library uh on the selected test data set and the quantum instance now quickly moving on to the next slide this is the high level architecture what i have just explained we did loaded the best sensor data set then did some pca uh imported i ibmq library i will loaded ip account and uh bring out some feature map content streams execution settings that we wanted combining all these we have created an instance of quantum support vector machine with the test data set a trend data set in the predefined feature map and then after training the model we got the classification result we did some analysis on the classification result and we come out with some metrics that i'll discuss in the later part of the presentation so moving on to the next slide uh this is uh overall view of how feature map actually works so feature map is a general general walk of the feature map is just to transform the classical data to the quantum data right so the algorithm works in the same way is the classical one the only idea behind the quantum kernel is that it tries to calculate the inner product but in the but on the quantum degree not on the classical data so in order to define a feature map there must be some entanglement as vascular has already applied now mentioned that entanglement is done between the qubits so as to measure the output of the one qubit the other other qubit is also being kept in the interaction so that we can build a circuit that would perform some operation and get the desired output so there is a different types of entanglement that is could provide it is one of it is linear one of it is full so linear renting element is something which is one qubit is uh connected interconnected to the next qubit and the next is connected to its next one and then in case of full entanglement all the qubits are interconnected but as we are running our algorithm with two qubits only so either we take it as a linear entanglement or a full entire element no nothing matters because because the two qubit has two grids are only there so the interconnection between interconnection or connection between those two will be a simple connection which we here you can see by the c not gate that we have used in the first image we can if we are using the multiple qubit machines or we if you are using going for the higher dimension this is for the purpose that if you want to go for the higher dimension we can explicitly mention a entangled map it can be a integral map is nothing but a dictionary which defines the interconnectivity of a source could be to the target qubit we can explicitly mention it in the feature map that we are using uh the feature map can be second order expansion powerless expansion that data on there in the skit library already so uh how this feature map actually was the meth behind all this can be seen in these equations so what it does is uh first in order to classify transform the data we need a unitary operator which is nothing but a breakup of three uh phase shift gate and two control not gate which is just to control node gate are just to create an entire element and unit tree gates are just to uh rotate the qubit into different phases uh and convert it to cluster from a classical layer to quantum data so the unitary gates that we are using is u1 gate that the metric of which is shown here which is a diagonal matrix of 1 and e to the power i lambda so here phi is a function that we are using set of functions that we are using to transform a data set from a lower dimensional sphere to higher dimensional space we are using this also to calculate the kernel matrices that we are going to uh get or where we are going to need in our determining our hyperplane so talking about the u uh unitary matrix that is the second equation this is the uh general equation which is uh working behind all the circuit right and we can explicitly define our circuit or the skillet library is already developed so much that it is already providing the secondary expansion as a function which inherently creates this circuit for us without delving into the equations and creating it manually so after creating the circuit like in in the third figure or in the second figure we can see that there are blue and red lines which which means that we are using only two qubit machines so two qubits uh so that only two lines will be active and the rest are not needed for us to perform our calculations so this is overall how feature map actually works how it transforms the data set how the kernel is calculated we'll see in the next slide so this is the math general match behind how the kernel is calculated so we can see these are the system of equations that is being sold to get the parameters for the hyperlink k is a kernel matrix that we already have seen that it is used in place of the inner product if the data set is transformed to a higher dimensional space so k is a kernel matrix which we need to determine and the only need of a quantum computer or only the quantum part that is being processed here in the usm part is the estimation of the kernel matrix only rest all the things can be done classically also so the kernel estimation which which is a time taking process is done uh in the quantum way and the rest all the processes is done in the classical way so k is the kernel matrix that is determined or the which already that we have created a circuit we have uh built in feature map which is second order expansion they do a task for us you know get the inner product of the data and why here represents the level of the classes it is predicted if it is a positive it is plus one it is negative it is minus one so only thing we need to return and in the uh hyperplane parameter that is b alpha so in the image you can see the simple linear equation which is w x plus b equal to 1 0 n minus 1 so w is similar here to alpha which is a a hyperplane parameter or we can call it as a slope so once our kernel matrix is calculated in the quantum fashion uh so we can use it we can use it to uh now calculate the hyperplane parameters hyperinfluent parameters are the only unknowns in the above equation and then once the hyperplane parameters are with us we can classify the new set of datums easily by uh inputting our suppose x0 is a data point which we need to classify so we can easily put it in the equation and we can apply the sigmoid function so it will return our output as a plus one and minus one so in order to calculate the um uh class for the new set of data again we need to calculate the distances of the new set of data to each and every point in the future space so again the quantum quantum part is involved here which is the calculation of the kernel matrix for the new set of data so we have seen that quantum phenomena is called to twice in this protocol first the kernel matrix is calculated for the uh set of datum that is already leveled and secondly the kernel matrix is calculated in the quantum fusion for the new set of datums and these two are the time taking process if we do it classically which should ideally be uh not taking time in the quantum version by uh making it parallely or making it kernel making use of the kernel any inner products in the feature space now using this equation we can uh easily classify the data in the plus positive class or the negative class so this is a high level of how what the math is going behind and what is the quantum part that is involved now coming to the implementation part we have uh uh simulated our four uh experiments which uh was on the breast cancer dataset one one was on the entire data set with all the features that is not relevant in this case because we can uh compare we can compare the result of the algorithms that are similar in nature right so what we did was we were using the quantum support vector machine with only two features so we already we also perform the support vector machine classification problem with two principal components and then we analyze the uh confusion matrix confusion matrix is essentially distribution of the data points as a true positive true negative false positive and false negative so we generated a quantum we generated a confusion matrix then we went uh to michael went to calculate the recall value for the svm with the classical data set um in in with two principal component and performed qsvm with two principal component on the simulator that is i think uh quantum 16 melbourne simulator and then we chose a quantum printer ibmq real real-time machine wherein we performed and we got a recall value of this to be 0.931 and 0.8 so quickly um quickly explaining recall will is nothing but of uh predict of how many actual cancer patients if you say how many are predicted uh they had cancer so in these type of problems that uh recall value is much more important than the precision or accuracy just because if we are not able to classify the cancer percentage as a cancer then it would uh be a life taking for that person right he will not go for the diagnosis or he will not go for the treatment so we mainly focused on our experiments and tried to calculate the recall value although at this point of time quantum machines are not quantum machines and we can also the quantum simulators are not that accurate that it could classify all the data once correctly uh being classifying it classifying it classically gives more accuracy the only advantage that we get here in the quantum or what i uh actually observed is we could get a runtime uh like we can reduce the runtime of the learning algorithm but the question that uh that ashes has already answered right uh that we did not see any uh runtime uh reduction while using the quantum super director machine at the view i have uh applied it few months back right it took around 20 to 30 30 minutes for calculating the or estimating the kernel and giving the predictions for the 20 state of data points only and we can run a classical svm of uh hundreds of data points or thousands of data points within fraction of a second so i didn't see any advantage in terms of the time reduction and we can't at this point or time expect the accuracy part just because it is in the development stage and and i think uh by the time we have applied the qsm algorithm and by the time uh by today uh when we see there must have been some certain improvement in the classification accuracy and recall value just because the i i have visited the biscuit library and i have seen that there is a lot more advancement in the feature map that they are providing and the feature map is the basic circuit basic way to create a circuit and the operation that we provide in that circuit ultimately decides that uh what will be the recall and accuracy so this is all from my side if any questions you might if i chime in again i'm wondering uh i'm not sure if i saw all the details on your kernel approach but in the classical algorithms you usually have to pick a class of kernels for the feature map in in this particular implementation are you claiming there's a level of generality with respect to the kernel where you don't have to tell it what type of kernel to use nice question but the answer to this is that the second order expansion that we are using here and the qscm estimator that we are using here this kit has already inbuilt uh what's a kernel function function that they already defined in the estimator itself uh it is rbf kernel at that point of time when we are using the second order expansion depending on the feature map that was uh that i was talking about that feature map decides that uh what uh what type of accuracy will you get and accuracy can be related in terms of the kernel function that we that you are talking about yeah so kernel function uh depends or i would say that kisket is already like introducing the kernel function itself in their code library so we need not to explicitly define it thank you so much [Music] talk i think it will take another 10 to 15 minutes then we'll hand it over to you mart okay thanks cool then so before going into the queue again i will first brief you all about what gans are so gans were found in like they were generated by young fellow in 2014 so it's basically a great advance in machine learning so they have numerous applications and one of the fanciest that i feel is a face generation so basically there is this page called that this person does not exist dot com which generates which shows your face every time and it seems real but it's not so yes it's one of the great applications of gans so basically gans is like a generative modeling approach using the deep learning methods so it basically has its like it's a game where there are two players both of them are neural networks one is generator one is discriminator and the goal is to generate data that resembles almost resembles the data that we are providing as a training set so the generator keeps generating the fake sample fake examples and both the real examples and the fake examples go into discriminator and discriminator tries to detect which one is real or which one is fake so there are these two neural networks one is generated and one is just discriminator and based on the loss functions that we calculate these are like the weights of both of these neural networks get up gets updated and eventually it like the process would stop when the generator is generating almost as real image as it is or as real example as it is in the real data so as shown uh in this architecture as well like the generator generates a sample the rear there are samples from the real world examples images and it goes into discriminator and finally detects and calculates the loss function and then eventually improves the model so this was the overview of can now let's look into a bit of mathematics involved in gan like the loss functions so bhaskar can we move on to the next slide yes so basically x is the real world examples the sample from our distribution let's call it p r or p real the g is the generator d is the discriminator of which the output that is 0 or 1 but it's a probability on which we say it's 1 for real or zero for fake so generator loss function it gets calculated using the output which discriminator gives that is g d of g of z so we calculate it using negative of the log for each sample and then we like basically average it out and for the discriminator loss function it's get it gets divided into two parts it like um the discriminator has to tell if the image or data from real distribution is real and the fake is fake so it has two components one for real and for fake and similar calculation gets done so basically the training gets over when these booth gets into a nash equilibrium since it's a game we can see it from the game theory perspective so basically one the generator is trying to fool the discriminator but the discriminator tries to like it tends to see if that generator it classified it as fake so yes so let's move into the queue and part so the queue can so basically there are three main components one is the real training data that we are providing one is discriminator and one is generator so based on that we can see different scenarios for implementing cuban so the first one which is stated here it's like it says data quantum discriminator quantum generator quantum so it's basically the best scenario for quantum where we can like actually perform with greater efficiency but since the systems are not that fault on tolerant so we were not like we are not able to implement that now so for now what we have implemented we have taken the classical data set the third third one that i'm talking about the discriminator is classical and the generator that is generating in quantum so we have implemented this one since it was the most practical application that we could think of at that point of time when we were applying so can we move on to the next one yes so basically this slide tells about the q-gan generator circuit so q gan is basically a parameterized quantum channel where we have like n qubit input state and n qubit output say it so the expression tells about the input input state uh of this where p j theta is the likelihood resulting occurrences of each of the states so the quantum gate that the quantum generator is implemented by variational quantum circuit so there are like this alternate layers of gates where each qubit that we have we apply the poly y gate which basically changes the y component by pi radians and between two of two of these units there is one like for this control z gate which is like entangled entangles two qubits so there is this truth table for this particular that when like there is one what do i say control bit and one is target bit so when one is zero and other one is zero it it gives the same output it won't change the target bit it will only change when both of them are one it will change the target cube target bit with like it it will flip it the it will flip the z component of it so yep uh so when we like uh when we start like with the initial state so that will like if that will mostly tell how much time it is gonna take in application since it's a process and after as many iterations if we start good we finish early so that is the scene and so how we did the optimization so basically the classic classical discriminator it's a standard neural network with the sigmoid activation functions so we have m samples from the real world data the actual data and we generate m samples from the generator so based on these both of these uh we train the discriminator and we calculate the generator loss and discriminator loss and since once if one gets minimized the other one gets maximized so it's basically a min max problem so we try and find a nash equilibrium or both of a point where both of these converge converge or like basically these are optimized based on the two parameters the phi and the theta for generator and discriminator respectively so yep so this is the architecture that we followed like uh from the quantum backend we pull kisket library and like for this particular experiment we generated for the real data we generated uh some thousand points from a log normal distribution and we put that into a quantum circuit so as to change since the generator is like and outputs the data in quantum format so for changing the format we push the real data into the quantum circuit which finally goes into a discriminator and from generated goes to disseminator and the output and then we evaluate the metric the loss function and eventually it trains the model again based on the number of epochs and number of iterations that we provide as input so for generating this output these like some of the graphs so we use some of the parameters like we defined number of qubits we defined the mu and sigma for this log normal distribution we define number of epochs we want we defined the batch size for a particular training we just we defined the number of samples that we want and based on the like bhaskar and amit already said there are already like predefined functions in the library that are there so we just put these parameters that we choose and we create one quantum instance and then eventually we can train a model so based on that we calculated the loss function for both generator in the first figure that you can see there's this generator loss function and the discriminator loss function so basically describe what discriminator trying to do discriminate is trying to maximize its uh loss function on the other hand generator is trying to minimize the loss function so eventually after the 3000 iterations the both of them reached some convergence if we like increase the number of iterations the model will definitely improve but based on the time constraint that we had so we continued till 3000 iterations and we saw these kind of scenarios which eventually like satisfies the theory behind and similarly similar case was with entropy so uh we calculate entropy by one formula which is like the ratio basically the ratio between uh generated and the real one so it like there uh the data which is getting generated uh should like more likely like the real data so the eventually the relative entropy decreased and the last one that we can see is the distribution one is actual normal and one that we got from our study so it was almost like imitating the distribution that we provided as the training set so this is what we have observed so we have like implemented for a one for one particular type of q gun that we can see but we can like basically what we can explore as many combinations from this uh real from the data type or discriminator type or generator type that we can see we can see different scenarios we can run and then we can like see the results and check like how things are getting better in which scenario and we can explore that way so we are in that phase uh this was for one of the types that we explored so yeah this was what was uh what we did in the q gan for implementation part so i think yeah that's pretty much like what we wanted to show uh so it was the entire presentation is all about what we explored and what we tried implementing and um like uh like and what we understood or learned so like as a as what i would say is like um anyway like we would definitely have a discussion many questions can be raised and we can also try solving them you would also like to hear about any experiences of someone like who implemented any of these kind of algorithms and um in a practical sense or in any in any sense and also we would also like to understand like if there are any suggestions about what else we can do to improve or what else uh in what in which direction which uh we can go in our research so these things we would be um like very excited to listen to but over to you most uh like you can take this forward the discussion the questions and everything okay um we have thank you everyone we have maybe like 10 15 minutes four questions also josh can do you want to introduce yourself because we kind of missed you in the in the first part right right um hi everyone i'm yosh i'm a third year at stanford uh studying math and physics uh potentially computer science in the future and yeah um i'm just really interested in uh quantum algorithms specifically like quantum computational complexity um and yeah thanks so do we have any questions we can open the floor for discussion maybe i will ask you guys a question yes can you hear me yes yes awesome so i mean this is a great exploration right it's very interesting to see kind of uh outside view on the q network and kind of what's advantageous advantages so uh i would say what do you guys find uh there's a sigmoid as a third party using this uh network what do you see here as an opportunity right and for you as a company and how can ibmq network help you uh kind of achieve it better right so given that you've done this experiments what's your feedback essentially and like in your roadmap for quantum exploration where do you see quantum business going how can we help you get there thank you uh alexey thank you for uh that sounded more uh like an offer as well as a question but uh frankly we are learners in this space uh analytics and data science is a very competitive space right and being ahead of the curve with always other companies and be able to implement or being able to identify solutions or techniques which can help us remain ahead of the competition that is what we are thinking right now and for that this is our experimentation start we want to ramp up on this uh thankfully as in from your side we saw interest and that gives us more uh enthusiasm to keep on working on this we had sort of uh put some brakes on it now we feel that yes we are going to work on this with some more funding internally as well as uh externally possibly uh from our side as you can see most of the uh problems that we see from our clients are real business problems and for those companies are still struggling or sorting to the traditional methodologies and algorithms to work on it as soon as this speedup comes in and with real time analysis becoming more popular the data growing very fast anything which helps us make or run algorithms at a very fast pace i think that's going to be something that we would want to learn we want to lamp up on that very fast we had only mentioned that with the wave we were not able to explore anything further with ibm queue we were able to explore a lot it's a good starting point for us we want to ramp up on this and to expand our r d capabilities as well machines is one from a hardware point of view a space where we can explore these algorithms and uh make them applicable in real business applications that's the space wherein we would certainly be looking forward to any help i think having it open source having it available for public to execute i think that's one of the already good initiatives from ibm queue we would want to enjoy the benefits of it in future as well awesome thank you thank you thank you and let me for experimentation so i have a training here uh for experimentation with star number qubits uh we have partnership so we can you know talk about it separately because you know there are much higher number qubits and i think for kind of interesting tests uh it would be interesting to try what you guys did at higher number of qubits great so i think don has a question right yes um i just put an article up there i found in physics aps physics about super conducting qubit that protects itself so that seems to be on the the hardware level uh where it's where everybody's heading you're in the algorithm level so what's the repeatability of your experiments i mean do you have a consistent error correction or does that slip between you know between the certain boundaries and then the error correction itself is is not totally stable yet to prove that the error correction really is error correction you know what i mean it's not that the results shift it's the actual error correction is actually more accurate because that seems to be what everybody's discussing is how close in the error correction can we get the prove that the error correction works that's all i keep seeing i mean you know maybe maybe it's me but that's all i keep seeing when i do when i did electronics that was the whole thing accuracy of how things were going wrong and how you had to trace things procedure so each time you did it you know you had 12 parts in here which one was actually doing and why was it doing it and the idea is you can replace all the parts okay well big deal you haven't proved anything it's just the idea can you get down to that one part that's actually got a hiccup in it because there was a malfunction in the in the manufacturing the chip we bought so you understood that the chips were actually cheap and we should go to a higher grade chips and we'll eliminate that entire problem because it gets in the engineering production phase that's when like ten thousand years ago so i think it's still relevant um but it just dawned on me i just keep looking everybody's talking you know so that's something to focus on like they this article i just found on superconducting qubits that protect itself seem to be getting the cubit itself in a position where it's actually corrected and it's stable so no other influences can actually throw it off so we have a better view of what's going on as we target you know getting an accurate assessment of how these things can go to a million cubits or 10 million qubits or whatever because the 60 you said 16 million or 16 qubits right which is really sophisticated at this point for what you guys are doing and just like you know i applaud that yeah you know and go with that but the idea is the next stage is they're trying to get to that that level i'm just saying this is some information maybe we can help you out certainly this helps yeah this would certainly help us uh further our research but frankly uh that was that question was way above my answer it's just something i'm as i'm kind of learning how this works i'm just putting there's things that stick i like this parachute i spotted this but somebody else figured this out when it came down i'm going like there's got to be some reference out there and all of a sudden i find out that some somebody in school who's much younger than i am and sharper and more you know more hungry found this out and figured out what they put in that message okay so there's those little things that when you get that little bit of excitement there you just want to kind of follow it which is what you guys are doing you know you have a lot of excitement for this i mean you really really want this to work and you really really seem to be putting the time into making sure that what you're doing is accurate so it justifies the work you're putting in there and even if you come to a conclusion where oops it doesn't it isn't but how do you do that stop but the next stage is incrementing more and more accurate correction in order to get where you need to go so i applaud that yeah yeah we are trying we don't understand everything we don't understand the technical technicalities a lot as well we are trying to work our way and be able to understand that sure thank you we will uh look into it it will certainly help us enhance our knowledge as well yeah i can like answer some part of it so as john said like so for this quantum computing for it to be an action so we need this qubit in action so eventually if we like there is some like external magnetic field that we provide for changing the direction of electron which is like it will tend to be it in the like the natural state or the low energy state so if we want to change the direction we'll have to provide certain external field for it so for maintaining that all the like i think that there is some research going on like how can we maintain it better otherwise there is one phenomena with this qubit that's called decoherence where a qubit loses its like power or what like the properties basically on which it's acting so yep so i saw that article so it basically says uh the title itself is a super conducting qubit that protects itself so it's basically i think i will go through the this article and i'll figure it out but mostly i think the mostly the fight is going on like how to achieve that state and how to keep that electron like for maximum time that we can right i found that when i was doing tech um doing electronics we had problems like i said with the chip we bought the wrong chips we needed to go towards a higher priced chip but they were trying to find them in bulk so the problem was we were having issues and it came down to looking at how these things are actually the lithography works and you get into the superstructure of the chip because every time an electron goes through a passage in a hole it's going to leave some kind of trail and it's going to bounce into something in the lattice structure i had not my head better books than i have today but of some really detailed books that basically showed lattice structures designs that basically as you see the electron could bounce through it as it bounces it vibrates and that vibrates turns into heat and the more that happens the faster you go which is why the cooling process is because it holds that more rigid so as it hits it it represents it's resilient as opposed to flying out and burning up so when you see your chips burn up on your computer your processor at one point remember remember overclocking you know when overclocking wasn't a big thing you know the gamers went out there want to overclock these processors and they basically were burning them out and they brought this to the manufacturers and they said hey we want to play these games faster than this how do we do this well then it gets down the clock speeds and crystals on the board and then they just had to basically process the material so it could actually handle the overclocking so now overclocking is just not it's not even mentioned anymore it's mentioned that as a fact it's a short-term goal for a higher level operation because they've already put in the manufacturing process so that's that's kind of all i was kind of that's where kind of where i was looking at things great oh hello uh yeah i just uh have a couple of questions uh so first i think bhaskar had like given about uh implementation of uh qm algorithms right uh probably uh qsvm or something uh i mean does uh running qml algorithms give any uh validating advantage quantum advantage other than runtime and if if not then what are the challenges that i mean uh like that are required and i mean for that i mean that had to be met actually yeah so as far as um the entire quantum space is uh known uh so runtime is the biggest advantage we get uh for any of the quantum machining algorithm or like any quantum algorithm except for maybe like as she was talking about the cryptocurrency or like encryption or decryption parts um like there we have an additional advantage of not just time but also the uncertainty as well but like when it comes to the quantum machine learning type of algorithms or any classical algorithms in a quantum state i see uh the time advantage is is the only huge win um yeah and like other challenges are like we have to create a lot of libraries around it and uh convert um like or build a lot of stuff with converting the data and build hardware so a lot of challenges around it but that's what i see any anything in this uh additional to add to this no at this point of time as in from whatever we have been able to run we have not been able to identify any advantage so far but that is not because the quantum theory doesn't work but that is more because of the current challenges of the way we are dealing with data we are still dealing with data in the classical sense right tabular columnar worst case scenario in a uh big data format but that's where it is i think the data part needs to align well to be able to run any qml algorithms so advantages will come but so far as in we have not seen that uh it performs worse than the classical algorithms well i mean across small industry sorry go ahead i don't want it right theoretically it should perform better yeah but practically it is not it's just that like we also have do not have uh great machines like maybe having better machines are having better uh built machines also that drastically would change the game anyway yeah mud you were saying something yeah so like across in the sea i think um there are two cases of quantum advantage proven quantum advantage and one is google and the other one is a photonic advantage from china um other than and they're not useful at all they're just toy problems so it is kind of naive to expect advantage right now i mean optimistic not naive i expected i mean tomorrow i wake up and when i do my research i expect to see advantage on my computer but uh realistically a quantum advantage on a useful problem is not is um is likely to happen um not tomorrow but soon and um it's not gonna like be an exponential uh speed up probably like we'll see small advantages through the hybrid algorithms where most of the work is done by classical devices and that small quantum component is doing the the the the thing it's optimized to do for and this brings us back to kevin's earlier question on um not on the hardware side but on the application side where do we expect to see advantage and i think it's a good question for everyone so can if everyone can like write their uh thoughts on like their their guess on it um on the chat i think it would be a great opportunity for us to share so the question is where do we think that sorry yeah go ahead yeah yeah yeah i guess there's something there's some work going on like you know i think there's some published work going on and quantum teleportation right i mean uh maybe uh does that have any relation with the quantum supremacy or quantum advantage that we are talking about i mean or that is something uh completely uh sorry um i couldn't hear the last question yeah so there's some work going on and like i think quantum teleportation and i think um you know you know so that does not have any relation with the quantum advantage that we are talking about you know one-time teleportation you know um i'm not sure what you're referring to so if but i mean all the work being done in the space is kind of related to quantum advantages like driving drive driving and research forward so yeah i wrote my answer for me it was random number generators uh real random number generators that's my thought that's um i think related to um what ashish was talking about in crypto crypto and security yeah um but yeah um i just have uh yeah sorry to interrupt you uh like i just have um one more uh like uh uh question to i think underwood like uh i mean can you leash like uh um overall like you know on a broader uh spectrum like what are the objectives of q gan actually like you know what are we wanting to achieve actually can you please do that like uh the main objective is to generate the data which resembles so it basically what it will do it will like it can improve whatever i have or like if i should say so basically the training data set that i am providing so it follows some distribution right so what my model is trying to do it will it it like it will try to estimate some density and it will the generator what it what generator will do it will like other than those samples it will try to generate more data from the same distribution so eventually my model will train for better so that is the idea behind creating it and other than this the gans can be used for enhancing like in the image site like if you work on that it can enhance it can boost the features of some image or i have also seen this implementation of can where if you simply provide one sketch it will colorize it or like it can generate some art artworks which are not there and which are like very good like since the gans have developed since the starting of gans like they have developed very much so people are actually using it so uh the site which i said no that person doesn't exist does not exist.com visit that one and see that image and try to like try to guess whether that image is real or not that image is the result of this one of the gans so okay it has i'd like to i'd like to learn more in the quantum format so we can use it for the algorithms or we can as said that we are able to we are trying to generate the data set which is which resembles some of the distributions like here in case in this case he has generated the data which resembles a long normal distribution right so we can generate some of the data which we want to train or classify as in case of quantum super director machine so we are using a classical data and then converting it to the quantum data so in in case of that in place of that we can directly use the qgan to generate the data which resembles the classical data that we are using to classify and we can further use it to give it input in the quantum circuit for quantum supervisor machine okay yeah thank you yeah okay i think we can start wrapping up also when i said there were two quantum proven quantum advantages i there's also one paper that came out last week so it's not i wouldn't call it a proven quantum advantage but it's it's a good uh good supporting paper for annealing advantage so um the wave just published uh last week and i'm gonna put it on the chat you should check it out if you're interested they claim a healing advantage for this very good classical method but it's not comprehensive against all classical methods um so but it's an empirical result that's interesting okay so thank thank you everyone um we will uh publish this video on uh our website which is quantum dot sv i also put that on the chat uh thank you all for coming thanks to sigmoid for taking the time for this very well prepared talk um we hope to see you all again uh next uh next month the last thursday of next month is that right alexi yeah that sounds good sounds like a plan thank you very much guys it's everybody looking forward thank you pleasure thank you everyone bye-bye for providing us with this [Music] bye