Shift AI 2020
April 21, 2020, Online, USA
Shift AI 2020
Request Q&A
Video
Shift AI 2020: Deep Learning in Intelligent Process Automation - Slater Victoroff (Indico Data)
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Add to favorites
229
I like 0
I dislike 0
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
  • Description
  • Transcript
  • Discussion

About speaker

Slater Victoroff
ДолжностьCTO at indico data

CTO of indico Data Solutions, turning raw data and image into human insight via machine learning.

View the profile

About the talk

RPA has exploded in recent years leading to never-before-seen levels of enterprise and desktop automation. As automation continues across the enterprise attention has turned from simple, deterministic processes to probabilistic document-driven processes now typically referred to as intelligent process automation (IPA). IPA has a number of new requirements and best practices distinct from those of RPA. Topics include human-in-the-loop learning, biases in model-chaining, and business-driven model efficacy assessment.

Share

Hello everybody. My name is Slater big draw and I'm really excited to be here at this virtual shift and I'm going to be talking to you about deep learning and intelligent process. So a lot of you have probably realized automation has been in the news a lot lately and people have a lot of very different takes on it since we're going to be talking about it wanted to go through a little context. Some people might be familiar with the sky Andrew Yang who the recent us Democratic presidential candidate making really big

waves at least here in the US by making automation kind of the central theme of his campaign saying that Automation in in very specifically this upcoming fourth wave what's being called the fourth Industrial Revolution? That's a lot of software automation is the central challenge facing us being kind of the world today. And anyone who's not addressing that is failing their constituents the same time, you've got people like Daniel, who are saying that actually the promises of automation global economy collapse and this is something that a lot of people might not be familiar with

but there are a lot of Economist out there that have basically Model did with an aging population we might not be able to keep sort of the growth of global communities in line at which could cause to a lot of really really big issues with Herman Denton versions. I'm so when you got these really kind of distinctive use on automation, right? You know, what I want to do is kind of cut through that a little bit and really show you what happened, right? You know, I've been working in this place for about seven years now. It's a pretty noose face. I wanted to share some some lessons that we got

from the front line. So quickly with you about automation change so drastically in the last couple years that has caused this to become for a really really big topic much more show than it had been until pretty recently and I want to start by talking about robotic process process automation RPA probably heard of our paid previously. This is a really really big industry. It's kind of exploded recently. Large companies like uipath automation anywhere blue prism but but kind of countless others, you know, if you're not

if you're not serve a finance person, you don't have to worry too much about this chart. Although it's saying is that this is a multibillion-dollar industry is growing, you know, upwards of 20% year-over-year and is projected to for you know, the next many many years to come but what is RPA you're not necessarily as familiar with it. This is kind of a quick quote from from uipath. Who as I mentioned previously is one of the real leaders in the space one of the largest companies over recent recent unicorn and it's it's something that is specifically designed to allow sort of

non-technical people. Right? So this is you know, if people are familiar with selenium is a kind of similar set of functionality to selenium, but it's putting that in the hands of a non-technical person and kind of enabling it on the desktop another way right in RPA you're going to have for the basic set of relatively, you know Road. Haddock recipes that you can take right in Cotton coffee things. You can open up different applications. And what you do is you combine those to put together a sort of a larger robotic, right? So for instance, you know, you open up Excel you copy some values you

put those into you know, you open up QuickBooks and you put the values in there, right? So that would be that would be an example of something that you might automated process in just some sort of basic terminology other sort of the concept of an unattended. You know, when you got this this full process running you say you haven't bought in production and you can have unintended bot or an unattended bot flies attended is, you know, a person sitting there watching the things run very possibly even might be that the full process is automated to make you run 1 Bots you type some things into

an interface to write you run second thought something like that ended right? It's it's in a rack. It just runs. It does its own thing vast majority. Sorry, can I switch screens a little better? Y'all of all our PO box today are attended box. Sorry, it looks like. Did not owe. I see the problem. Sorry about that everyone. Okay now. Sorry, for some reason it will not actually switch over in obs. Sorry everyone. I'm trying to get the sizing of these to write. But it won't actually switch over.

You know what me. Thought you were going to do this kind of an exciting Moore LifeWay. Okay. Alright, so it looks like that that actually has been successful apologies everyone for the for the technical issues. But now hopefully you can see my slides a little bit bigger or this will be a little bit more easy to follow. Right. So as I was kind of saying he has exploded recently and for these kinds of simpler automation task has been really really successful. What this is really giving rise to is. This

is our kind of further attempt automation right for this new field called intelligent process Automation and that's what we really want to talk about today IPA in this is kind of a definition of IPA from Mackenzie who is really one of the leaders in the space. It's really the idea of trying to leverage both the existing sort of business process tools, right NF things like RPA as well as primarily AI tools and using those in combination to kind of tackle repetitive replicable and routine tasks for today are kind of done by

people largely and learn to do them faster with more consistency you're better accuracy, but we'll talk a little bit about exactly how is anthrax long-term. So these are a couple example use cases if you're still kind of struggling to figure out. Okay, you know, what is IPA? What is RPA? These are sort of canonical examples write something like contractor review wear. If not just about getting a number from Excel and posting in QuickBooks. Right? But there's actually some interpretation to have to happen. That's why this is intelligent process automation as opposed to gather your things

like invoice processing corporate in boxes where you have to Route incoming emails rape really anything where you got a relatively simplistic process that today, you know, surprisingly enough. These are done entirely manually write someone looks at every incoming email and figures out who has to respond to that referral form write someone brings out the PDF of the form. They made you leave key in the values for each particular piece of those types of prophecies. Sota kind of recap rape RPA you can say things like coffee the

values into QuickBooks need selfie, right? You can say, you know, if you got Salesforce and your Excel sheet, he didn't get up to other see around you can say, okay. I'm going to enter the information in my IRP application and then the arcade. We'll go and enter that into several applications right see you on the other hand write this is anything where you got some interpretations of intelligence. That's that's going on in the process, right? So checking the data protection in the contract, you know, you have to read through the contract. The parts that you have to understand what is a data

protection and in a lot of cases actually do some interpretation of that in your model. We're checking all of the documents in a loan approval are are in good order, right? So when you look at a alone and when someone applies for a loan and when they get a potentially subsequent approval, there's a whole set of documents and it actually is kind of related to the Presentation earlier you have to check that all of the documents kind of agreed with each other right when someone sends me to pay stub and send them an application that says

how much money they're running and also sends in some ID, right? You have to check that those are actually all for the same person, right? And when you look at the standard, you know, it's slow and it's actually kind of surprisingly error-prone. And when we wonder why is I feel so interesting write it if we look at the kind of successful automation initiatives out there, right? So the entire world, you know RPA and AI use cases who's actually getting its production in and who's

actually getting for a lot of value from these received that folks using again. That's kind of really really heavily MLB's techniques here wood processing in particular for reasons that you'll see in in just a second or so while I see a is still relatively nascent right? It's still kind of a new space that's not nearly the size of RPA in terms of the full automation landscape. It's growing very very quickly and especially kind of top organizations that have been a lot more successful in automation are investing in

this kind of disproportionately Mackenzie calls. Flush the automation imperative r i a n Denver folks that have been spending a lot of time in their price offer certainly in the last couple of years, you know, you seen automation centers of excellence show up, you know across the entire Fortune 500 and then truly is this has gone through Surfer really really categoric C shift, but I think what people you know, what's not necessarily has obvious is that there is this disproportionate focused on start of

these machine learning driven process automation. So right, you know practically what is what the reality of the space right, you know what happened and what kind of actually going on the modeling of this right and what we talked about intelligently automating a process. Like I said, these are typically deep but we need to look at what is even, you know a pretty simple problem here. Right? Let's say I want to kind of extracted this this $5 that was really kind of classic IPA use case. The problem is that I got all sorts of the

visual structure here, right? You know, I have to understand that this basic write this 504 doesn't just apply to that but there's also the header for base right? I have to understand not just that this read order right? I have to go December 28th 2019. I can't just read this through like a line, right? That would be December 28th, December 29th earnings per share 2019 2018 issue is actually really really tough and you can kind of recognized as very hard for computer to automatically figure that out and then we realize that even if I get that block appropriately that's not even

the whole issue because I also have to understand that there's this had a relationship right three months ended to this date free simple example things get a lot more complex than this but we start to recognize really quickly. A lot of these IPA processing is don't fit neatly into the traditional set of problems. This is not really classification past. It's sort of a named entity recognition tasks. The issue is that when you think about traditional modeling approaches right there, they're really heavily stove pipe by what kind of day that you're dealing with, you

know, everyone's familiar with sort of the classic computer vision image common lately for some reason all of texts Network architecture nowadays are named after Muppets. So, you know Elmo Bert and Ernie are legitimately actually some of the most popular for the best-performing and models in past several years audio there hasn't been quite as much but you know, where is Weidman in the point is still primarily just that all of these approaches are completely separate and we think about something like this. This is clearly not just a

text problem, right the squealing not just an image problem, but what's fundamentally Difficult about this is that you have to somehow How to combine those two strings and there aren't really good data sets for this stand in front of machine learning perspective is a really challenging problem in fractus what people end up doing is actually just focusing on text. Right? So the idea is rather than actually trying to model the complex and it's here that is kind of still a very active area research something if you want to learn more about look into so rich and beddings and model surgery.

That's really a field Focus largely on the sort of problem is what you do functionally if you just run everything for birth, right is this is this is not really you news I think for a lot of people ENFP space but what you do is you take an input like this you run it through some sort of traditional off-the-shelf OCR and then you call the hole for the extraction task as just a pure, you know, kind of a custom need any recognition. So challenging right? There's absolutely still more work to be done to your the problem is that you know, if Try to really tackle this

problem head-on. I will never end up making progress It's just it's too difficult to difficult problem. The second piece is when we actually talked about automation automation is largely metaphorical when you look in practice and getting these into production most things that you're going to fall into what I call augmentation. I'm just going to explain what I mean by augmentation acceleration true automation, which is actually quite rare augmentation. Is that a human is still looking at every document still processing every document but what happens is that you have sort of the

computer system whose job is to look at the document and provide information to the human that will allow them to do their job more quickly, you know, sometimes it's closed loop. So there that humans feedback gets fed back into the model thing from acceleration where the lady is a computer will selectively decide to sort of bring a human into the process if it's not able to automatically you do recognize that is extracted. The right thing or the conference is Quite sufficient to get executed right? This is where you have your documents. They're hitting the computer, right? You know,

they're actually, you know, your model is extracting whatever value to a parole and there's no human intervention so over this is rare for a couple of reasons, but one of the really big issues actually is that you are not legally allowed to have an architectural like this for gdpr reasons. I'll just go through this super quickly lose a lot of debate about exactly what people mean by explain ability and what is gdpr mean by explainability the most important clarification they've given recently is recital 71 that says fundamentally explainability can only be met by comprehensive testing and any

human who is subjected to an automated decision must be able to do in practice. What you need to be able to do is actually while this might happen in automated way anyone who's evicted. So if someone gets their loan right by an automated algorithm Still fundamentally need a person in the slow sets it when that rejected a person can manually examine it in a full automation scenario there still people involved. And so when I say automating is largely minute, that's really what I mean the final and maybe the most important lesson that I want to touch on is that the existing process is often

assumed to be really really good but as we've kind of realized these are heavily heavily biased process. He's right there ones that for some reason, you know, when you want to automated solution, it's very important that you measure it but for whatever reason when it comes to only you know, what pure human solution no one really does decide to measure it into the thing that people don't realize is that when we think about these prophecies we think about a loan approval we think about reading the date of birth of the rights. We would never make the mistake, right? You know, we're sort of

this is very professional right? We think of ourselves at our best when we just had a cup of coffee in practice know what's happening is that the people who are actually doing these processes right are doing M468 10 hours straight day after day mistakes accumulate because humans aren't meant to do that kind of routine repetitive work for long periods of time. I actually don't quite have time for this analogy else get past it just to say that these errors show up in production constantly and many people are surprised to learn that they may have sort of his largest 10% of

their entire body of loans might actually have issues according to their own internal process, right? You know, someone may have forgotten their identification or they're just might be fundamentally inconsistent documentation. They don't even know because because these kinds of mistakes or so so common, so just in conclusion to kind of wrap up right intelligent process automation, it's it's super interesting space and it's not something that you should really think of his being far off. Any more of these applications are going into production everyday, you know, the fourth Industrial

Revolution is is well underway spite of the fact that deep learning is moving so quickly and the models here and sort of the approaches have gotten drastically better in the last couple of years. It's still a very very nice in space right it Still very new it's not something that has were very comprehensive set of best practices. And even though people talk about intelligent process automation really readily as I said before automation is extremely rare, right? Is it certainly less than 5% of all use cases out their true automation for for very very good reason

and finally humans computers cooperate much better than they compete when you look at the models of this are actually really being successful right? It's not full automation, right? It's not models reaching Human Performance and then you don't have people do it anymore. Right? Just like the original Industrial Revolution that analogy is becoming a lot more apps, right? Cuz it's humans and computers working together to complete the process much more effectively in a more consistent the last last one. That's my presentation, and I think we've got a couple questions.

Cackle comments for the website

Buy this talk

Access to the talk “Shift AI 2020: Deep Learning in Intelligent Process Automation - Slater Victoroff (Indico Data)”
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free

Ticket

Get access to all videos “Shift AI 2020”
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Ticket

Interested in topic “Artificial Intelligence and Machine Learning”?

You might be interested in videos from this event

April 21, 2020
Online
14
4.57 K
ai, ai responsibly, automation, data mining, deep learning, graph deep learning, machine learning

Similar talks

Ehsan Yousefzadeh
Product Manager at Amazon
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Boris Cergol
Head of AI at Comtrade Digital Services
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free

Buy this video

Video

Access to the talk “Shift AI 2020: Deep Learning in Intelligent Process Automation - Slater Victoroff (Indico Data)”
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free

Conference Cast

With ConferenceCast.tv, you get access to our library of the world's best conference talks.

Conference Cast
613 conferences
24811 speakers
9162 hours of content