INTRODUCTION TO REMOTE DATA SCIENCE | ANDREW TRASK

Key Takeaways

[Music] hello everyone and welcome to this talk what is an introduction to remote data science my name is andrew drask and the next few minutes we're going to be walking through three main steps so first we're going to talk about motivations for remote data science so first is going to be some motivation for remote data science we're going to see a little bit about how it works and we're finished by talking about a new opportunity for you to learn more about this exciting technology so first let's talk a little about the motivations first setup for this is really about how you know our ability in society to answer important questions is limited by the amount of data that we have access to because that data simply is not at our disposal in another country or another department and this leads us to kind of the core premise of this talk which is that solutions to really important problems already exist but we just can't access those solutions because we can't access the training data and the machine learning data that we need to be able to solve them so there's a wide variety of examples that are like this but we're going to drill down into just a couple so first one we'll talk about is breast cancer unfortunately around one in eight women encounter breast cancer at some point in their lives around the world and this leads to as much as 700 000 of deaths per year from breast cancer and one of the main contributing factors to to this death rate is that we have a really difficult time detecting it meaning when you go in for a mammogram and a radiologist takes a look at the scan many places have as high as a one and four false negative rate meaning that one for women who go in who have cancer are told that they have a clean bill of health then they wait two more years for another scan and the cancer is really advanced by the time they find out about it breast cancer classification should be a relatively straightforward image classification problem which deep learning machine learning is really good at but unfortunately machine machine learning models even the best ones that are put into production are even worse than typical human-based screening programs with a one and three false positive and false negative rate being common and this is because uh it's my opinion that the largest data set that we have to work with is only three million images and in you know deep learning machine learning communities we know that this is not a very large data set and this is despite the fact that around 8 billion images have been produced in the last 10 years alone so we have this huge latent resource of images to work with but we're only working with tiny slices of it as a result machine learning model's it not actually solving this problem despite the fact that they probably could so this brings us back to our general theme of you know how solutions exist to important problems but access to data for those solutions is limited and we'll talk about one more use case real quick before we talk about the actual technology this use case comes from the united nations privacy enhanced technology lab or the pet lab which is a a new body that is piloting a lot of the technology that you're going to talk about today specifically they're looking at a trade related use case so you may not be familiar but you know canada and the united states for example each measure the amount of goods that are going over the canadian u.s border but often these things don't match so usually they match just like they do here but sometimes goods can disappear right where canada has record of goods that are supposedly going to the united states but the united states doesn't have any record of having received them it turns out this is a really important thing for for nations to be able to measure and so the new u.n pet lab project is is working with five different um member nations to be able to collaborate together on their shared data resources to understand how goods are flowing without having to centralize those data assets into into one location right and again this is all just about the same core premise where there are solutions solutions to really important problems that already are exist right in theory but access to the data necessary for those solutions is is too limited for us to to accomplish them now in this talk so far you might be wondering why not just centralize the data why not just just put all the data in one place that you need for say breast cancer or monitoring global trade um to to solve these really important problems and and there's a host of different reasons that we could consider why this might be the case but i want to focus on one that is it's simply it too inconvenient um and and just anecdotally you know if you want to solve this breast cancer classification problem that i mentioned and and you're going to require data from a thousand different hospitals that's a thousand different institutions that you have to get on the phone with that you have to talk to their legal department they have to talk to your legal department you have to go back and forth and back and forth and you have to manage the risk and they're going to ask where you put the data like who are you you know things like background checks and it's just such an enormously complex and time-consuming challenge that it's just it's just intractable even if you have lots of money and willpower no one has overcome this challenge thus far and and this reminds me of how sort of public data used to be in the sense of you know before the the internet if you wanted to know about a business or prices you would order airline tickets you had to get on the phone right you had to actually call a business and talk to a real person and ask them and and and today the private data industry is still a lot like that and what we're going to be talking about in the next few minutes is really the same solution to kind of the same problem which is we should be able to take private data put it into a server and other people should be able to answer questions using that data without us having to be involved as data owners at all inversely as a data scientist i should be able to just access the ability to answer questions using important data in the same level of ease that i currently access the internet so this is the field of remote data science and this is really what remote data science is all about so you have an organization that has data you have a data scientist who wants to be able to work with it and they remotely access this server to answer their question instead of downloading the data set itself and this isn't all about data scientists accessing individual institutions it's actually about data scientists accessing a whole internet of data that is that is sensitive and private that they never actually obtain a copy of but they can still use to answer important questions so in the next few minutes now i want to transition to how this technology actually works this is going to be a very short sample of what this technology is actually like and if you if you sort of like what you see and want to learn more then you can check out the educational resources that we'll talk about at the end of this talk so this is going to come from a library called pisces and piscep allows you to have the additional functionality in partnership with with tools like pytorch to be able to study data that's located a different machine so let's start at the top with a simple experiment where we're going to to to locate data and do a little data science project so i import sift first thing i need to do is is source a network so a network is just a registry of where private data assets exist think of it like sort of sort of google search but for private data instead of publix i take a registry i connect to that registry and inside that registry are what's called domains so domains much like a website domain are just different servers that exist on the internet that are holding private data right so let's say i choose two of these domains and and now i get a variable so we'll say canada in the united states that that allows me to talk to those domains right and to be able to interact with them so i connect to both these domains and this brings us to sort of the analysis so now i've got access to data in canada and i've got access to data in the united states right so two different locations so the first thing i need to do is figure out what my privacy budget is and the privacy budget is that is the the limitation on the amount of information i'm going to be able to pull out of these servers automatically so it looks like i've got 200 epsilon privacy budgets from canada and i've got 200 epsilon privacy budget from the united states sort of to work with right and if i run out of privacy budget i can't pull any more data out can't pull any more sort of information out that's that's what keeps the data safe so i'm going to select a couple data sets so so see the data sets would allow me to list them i'm going to select number 32 looks like i've got you know the the amount of imports and exports of dairy that the united states knows about to canada and canada united states inversely we have the same going going the other direction and now that i've got these two data sets or i've got pointers to these two data sets variables that are pointing to data in two different countries right now i'm gonna do an analytic so i've got this pointer on the left usa exports to canada i've got the pointer on the right candidate imports from usa these are each pointing to tensors that are in different servers on different parts of the world it should be impossible for me to subtract them by each other right the data is on two different machines but what happens right here behind the scenes is just secure mpc shares using using an input privacy technique called secure multi-party computation and this returns to me a pointer to these encrypted shares at both locations right this means that i perform this subtraction but no data has actually left either institution instead an encrypted protocol was used to to be able to actually calculate the result and maintain the result in an encrypted state with shares edge dislocations and finally in order for me to download the result what i need to do is add some differential privacy noise to the result so that the automatic privacy budgeting accounting mechanism goes uh so you had 200 privacy budget left so you just add this amount of noise ah looks like i'm going to let you download this result right and i even have a little bit of privacy budget left over and what i hope you see in a sort of end-to-end float is that sort of in the span of one notebook i can go out on the internet look for some data find some data perform computation against that data even if it's it's you know data located on different machines and then download my result without having to get on the phone with anybody without having to you know do data partnerships and most importantly without having to actually download a copy of the data set and if you'd like to learn more about this i'm super excited to announce that with the general support of pytorch the open mind community is releasing today a brand new course teaching you how to participate with these types of technologies both how to deploy data into a server and advertise it on these types of networks as well as how to participate as a data scientist in and this will all prepare you for a forthcoming united nations certification on remote data science um so check out these courses please go to courses.openmind.org you can register for free just like 8 000 other people already have thank you so much for your time attention i hope you have a great pie towards dev day