Latest / Agentic Conversations (formally mlops.community) / Build a Culture of ML Testing and Model Quality // Mohamed Elgendy // MLOps Coffee Sessions #76

Transcript

Build a Culture of ML Testing and Model Quality // Mohamed Elgendy // MLOps Coffee Sessions #76 · Agentic Conversations (formally mlops.community)

0:00
Welcome back to another and The Lofts Community podcast.
0:03
This is a coffee session with no coffee.
0:05
No. Adam.
0:06
And I are here to talk to Mohammad today about testing
0:12
Centric development, not test-driven development.
0:15
Adam how was it for you, man? We just got done talking to him.
0:19
I thought it was awesome. I really appreciate his
0:20
Viewpoint. Would you think?
0:23
Yeah, really good. Like I'd said to him off line
0:25
there. There's lots of stuff about the
0:28
kind of workflow that I've been thinking about recently and it's
0:32
it's probably actually speaks volumes that even I'd missed
0:35
some of the real subtle stuff that he's talking about the
0:37
holistic view thing was really, really cool.
0:39
I really like that. Yeah, yeah.
0:41
So for those of you that do not know who Muhammad is and it's
0:45
probably hard to know who Muhammad is.
0:47
If I'm not saying his last name, it's Mohamed El gandi.
0:51
Hopefully I'm pronouncing that correctly.
0:52
Actually I didn't ask him how his name was pronounced, but he
0:56
is X Amazon extra Leo. He's done some really cool
1:00
stuff. I mean he gives talks at
1:02
Amazon's Devcon Google I/O. He's got he passed toxic
1:08
O'Reilly's. AI conference.
1:10
The guy's pretty notable in the space and he's also written a
1:12
book with Manning called Deep learning for vision systems if
1:16
you want to. If you're really enjoying this
1:19
conversation, then we'll leave a link below.
1:22
So for where you can pick up that book but currently he went
1:26
off and he did. What most people have been doing
1:29
recently is starting their own companies and he's got this
1:34
company around testing and the company is called quote.
1:39
Wait, let me get this straight, it's kalenna.
1:41
I was about to say koala, but it is kalenna.
1:45
And and so it's all around testing and I think this idea of
1:51
testing It's not the first time we've talked about it.
1:54
It is very important though. We're starting to see it more
1:58
and more, you're starting to see things like he mentioned in the
2:01
talk, like the evaluation store and you're starting to see
2:05
monitoring companies, start to shift and pivot a little bit.
2:10
I saw just the other day a monitoring company come out with
2:14
a testing framework and open source testing framework,
2:17
because it's almost like it goes hand in hand.
2:21
Yeah, for me again, coming into the tool.
2:24
I was very much primed to think of things like, yeah, software
2:29
testing, unit testing and things like also data quality reports
2:33
and the tools out there. But this year this is different
2:36
and the holistic kind of end-to-end view, I think is
2:40
really important. The other big takeaway for me is
2:42
to start thinking about maybe Co defying thoughts, run failure
2:46
modes of stuff like that. That was a really cool point.
2:48
So, well worth getting into the talk.
2:50
Yeah. And so for me, I think I really
2:53
enjoyed when you're looking at the documentation that is
2:58
inherent with testing, like the more you test, the more that is
3:02
just a side effect that you have a paper trail or you have these
3:05
artifacts which can become documentation in a way.
3:09
And so how that can help with this, what do they call it?
3:12
Like, the bus problem that if somebody at your company gets
3:15
hit by a bus, then all of that information.
3:17
Hopefully is not in their head. It is Demented somewhere.
3:22
And also, as I mentioned in the conversation, how that can help
3:28
for people who are having to jump through the Hoops of
3:31
Regulation, as we both know, because you're in, you're in the
3:38
the UK but it doesn't officially fall under the EU regulations.
3:42
It doesn't anymore, does it? No one will but we seem to have
3:46
a strong. But yeah, we Center this this
3:50
practice of Emily regulations. Anyway.
3:52
So that's a sore subject. We don't talk about that.
3:55
She just yeah, you do a GitHub. Pull request, huh.
3:58
And or you do it? No.
4:00
Sorry. You do a clone.
4:02
You do a GitHub clone and then you take the regulations.
4:06
Exactly said it's. So yeah.
4:08
This is our conversation with Mo and I'm Demetrius Brinkman.
4:12
I'm here with Adam stroke of whose name apparently means
4:17
fried prawn in Swedish as he told us last Last time he
4:21
guest-hosted. I knew it was something
4:26
something around there. So anyway, thanks again, Adam
4:31
for coming on and joining me for this one.
4:33
It's always a pleasure to have you.
4:34
I love the way that you ask questions and your mind thinks
4:37
for everyone else out there, I hope you enjoy.
4:43
All right guys we're here. We're ready to talk.
4:47
I'm super stoked to talk with you Mo.
4:49
I know that this has been Long time coming.
4:51
We chatted what like months ago and we wanted to go over testing
4:56
and testing has become something a bit of important topic that is
5:03
getting the light that it deserves shown on it.
5:07
And so today I want to dive into it with you.
5:10
I know you're doing cool stuff with your company around testing
5:14
and so maybe you can start with what you're doing at your
5:18
company and then we'll dive into more testing questions as we go
5:22
on. Yeah, absolutely.
5:26
I think it would make much sense to try to take one step back
5:29
into why we're doing ml testing. Well the story will give her a
5:35
lot of context on what we're doing.
5:38
So my background has been in the ml side before and I'm currently
5:42
the CEO and co-founder of Corona is more of a business side, kind
5:45
of role. But my role has been always on
5:47
the engineering side of my career.
5:50
I've been building and managing ml teams in different
5:53
Environments. And one of the places that we
5:56
worked for who worked for a a early start early stage startup
6:01
in the CV space building algorithms to detect defects in
6:05
x-ray. Machines, we're selling products
6:06
to the security and defense face the deal deities.
6:09
All the kind of like another customers at this level, we went
6:13
into it there where we're thinking about, okay?
6:15
Let's let's do a build a model, a second detector.
6:19
And we say 98% and whatever kind of metric the we're We're
6:23
producing on realize that this doesn't doesn't fit the bill for
6:27
these kind of applications similar to a lot of mission
6:31
critical applications. You can't go on like there's 99
6:34
percent chance that we're going to take the gun but one percent
6:37
chance. There's a gun going up there on
6:39
the plane. So we at this time we start
6:42
thinking about how the product and the the entire goal of the
6:46
company is affected by my the testing problem.
6:49
So we would push by the market and by our customers to think
6:53
about the Process and think about our current model training
6:57
process and how we can stratify their our work our data into
7:04
real scenarios, real product objective scenarios that we can
7:08
go to customers and tell them. If it's a gun, we detect a gun
7:11
in these scenarios with this kind of confidence and other
7:15
levels and so on. So we can have iterative
7:17
improvements so on. So at this point we did well
7:21
with that tell time and then we either become You got Acquired
7:24
and then I was writing the, my book and computer vision, it's
7:28
called Deep learning for vision systems, it was published by
7:31
Manning and I start thinking about how do we actually solve
7:34
this pain that I've always been lacking in that like a manager
7:38
of the team or leading the remote teams in Margaret
7:41
organizations? And then I've always been asked
7:43
all these hard questions like a Muhammad why did we fail in this
7:47
scenario or are ready to deploy? Are we ready for a demo?
7:50
And we always do our best in the statistical.
7:53
Methods that we have today to have some confidence in our
7:56
products but in my heart it's always been, I hope so.
8:00
Right. And we work with my team's, I
8:02
can't build a systematic process around that, so that was a pain
8:06
in my career and we started thinking about when I was
8:09
writing the book I felt like let's dedicate a chapter
8:13
research it. And write something about this
8:15
topic. I worked with a, with a few
8:17
researchers who are working with me on the book, and then we came
8:20
up with a rigorous systematic and mold testing framework.
8:24
And then we are likely you realize that it will never be
8:26
used. It's just very tedious and
8:28
time-consuming to be done without proper tooling.
8:31
So this is what led to our current company.
8:35
Now Kelowna, we started working on, we quit our jobs, on the
8:39
start working on, including a full time, so the back to your
8:43
questions. What are we doing here?
8:45
We're building a machine learning testing and validation
8:48
platform and the goal is goes beyond the kasinga's as we
8:52
realized as Working with our customers and our personal
8:55
experience, we found that that actually the entire lifecycle of
9:00
the model development revolves around training.
9:03
But we defined that, if you shift that into the test Centric
9:08
development things, get much better.
9:10
Every Upstream process is much more efficient in the terms of
9:14
data training and experimentation and the
9:16
downstream processing production is like much much less
9:20
production problems. So we realize is that the vision
9:23
here? It grew from just testing into
9:25
building a culture of ml quality in the organization.
9:28
It's not a one-time heroic effort.
9:30
It's more of a team culture to guarantee that's kind of quality
9:36
level improves my time and being able to measure it and
9:39
communicate it with everybody. The site is really interesting
9:45
and it brings me to a point that I think many people might ask
9:49
you actually, if you go through this journey and you've probably
9:52
got the other four but you use. All right, ml testing and
9:56
obviously the big there's a big difference in unit testing
9:58
integration testing all that, my mind immediately goes to data
10:02
quality and thing and for tooling, so immediately go to
10:06
toiling, like TD, a DQ great expectations, so could you
10:11
elaborate a little bit? And why they're different to
10:14
what you're doing. Yeah.
10:16
I mean we can we can look at the testing node in the workflow as
10:22
a node or we can look at it as a work flow on its own and build
10:26
that training development on top of it which what we did.
10:29
If you look at testing let's actually not.
10:31
Look at testing, let's look at model quality model, quality is
10:34
a function of the model and the test data and then we all our
10:39
work is focused on the model. This is getting much easier with
10:42
a lot of tools, making the model piece and the training piece
10:44
much easier. And then, Jessica is usually an
10:48
afterthought or less managed, basically, Benchmark data set
10:52
that you keep adding to and then that doesn't end that dislike.
10:55
If you want to build a kicking theory, yes, the the testing
10:59
peace can be a tool or two. But if you if you want to look
11:03
at the testing piece as managing the test case manager and then
11:07
running the emphasis getting reports comparing models
11:11
compelling and you have an eval store that says all these
11:14
results and being able to To build analytics and services on
11:17
top of it and Reporting Services becomes an entire whole new
11:21
game, then we thought about it about, okay?
11:23
How do we start testing, how do you start solving the testing
11:25
problem. We thought about.
11:27
Okay, let's build that test case manager.
11:29
But again, how you're going to deal with the results and then
11:31
the test case manager is is a whole.
11:33
I keep looking at that test data.
11:35
Especially in the vision problem is much much harder.
11:38
If you want to start a fire, your data into scenarios, What
11:41
other option do we have? Now?
11:42
Oven labeling is right? And that is very expensive and
11:46
time-consuming. Now you're the reason he was,
11:49
like, you have, let's say, hundreds of thousands of images
11:51
in your benchmark data set, you label them as the way, the
11:54
understand your domain now and you learn more about your
11:57
domain. You go ahead and relabel all of
11:59
this and it doesn't end. So we realize that we have to
12:02
fix the problem from here. So we built a test case, Studio
12:05
allows you to study file, you create new scenarios within
12:08
minutes as opposed to weeks. And so on, then we thought,
12:11
okay. The end of it now but we need to
12:13
manage these test cases and version them and be able be able
12:16
to. We don't make sure that because
12:18
soon enough when this is the process is easy which what we
12:21
want your test data grows from 200 to a million to 10 million,
12:25
to 30 million and your test cases grow from tens to hundreds
12:28
and thousands. It becomes really hard to manage
12:31
with our you know and it development environment
12:34
notebooks and so on. So we found no other way other
12:37
than building a full platform to manage this whole thing just to
12:40
make sure that the End-to-end process finishes in minutes as
12:44
opposed to weeks. If you have to put in different
12:46
tunics, I can serve just a short answer to your question, Adam.
12:49
I can say, yeah, we can put around some to links and some
12:53
like, see GOI actually does something like this software
12:56
development side and the testing side we can put in around some
13:00
tools, which what we did in into coming into this process.
13:03
But the end of the day, if you want to build model development
13:07
testing, it has to be a full platform to take the process.
13:10
And to end is like you said, Before this call Engineers hate,
13:13
to be told that's more. They feel like it's always funny
13:16
but we have to deal with the whole thing to make sure that
13:20
it's Engineers like it. So if I'm understanding this
13:24
correctly it's really trying to take more of a holistic view at
13:29
testing and was recognized from a layer above almost that we
13:36
need to bring testing into every piece of what we're doing.
13:42
Absolutely. And on top of that is you can
13:45
build a process. On paper, I worked a lot in
13:49
change management process, change management and you can
13:51
put a process in paper but having a knowledge base and
13:56
organization that builds on top of it is the actual the gold
14:00
mine here. That that make sure that the
14:02
process that product improves month after month year after
14:05
year, if you have a new engineer joining the team today, they
14:09
will ask, you have to ask a lot of questions as if you're
14:11
building a contractor. How is our models?
14:13
Are our what other? What are we what?
14:16
That's how we do? Your tents, all these kind of
14:18
things. Whereas, if you go into the
14:21
classroom and just look at the list of test cases, I look at
14:23
the history of test. This is so even when you're
14:25
looking at your manager you cannot film your team's
14:29
performance. Nothing here about development.
14:31
You just looking at the team and your product.
14:33
So instead of looking at the team as we have a model went
14:36
from 1991 and so on, we want you to look at, we have features to
14:41
detect a car, a train and it has improved this way and regress,
14:45
this way, we have a feature that is Zach's a car.
14:47
Our with certain level of occlusion.
14:50
So under strain on top view and side view.
14:51
So you have all these history on the knowledge base that's built
14:54
here with them for five months, you have a lot of informations
14:58
on a lot of analytics on how to identify failure modes,
15:02
automatically and how to improve the model as an.
15:06
Except what to do, is it adding more data at tuning more?
15:09
All the stuff happens when you have this knowledge base, Yeah,
15:13
so let's double click on that, for a minute because I know that
15:16
we had spoke about the idea of testing as basically bolstering
15:23
the documentation. And it feels like, that's what
15:25
you're talking about here is like, because you have testing
15:29
and because you're making sure that this knowledge is not
15:33
siloed in one or two Engineers Minds, then you can have
15:38
something that is more robust and anyone In come in and get up
15:44
to speed much quicker than if you didn't have this yet.
15:48
So yeah. Can you go into that a little
15:50
bit more? Because I'm not sure, I'm fully
15:52
understanding it. Yeah, absolutely.
15:55
So let's go again with the simple example, like a car
15:58
detector. First, you go ahead.
16:01
And you say, I want to build a car collector and then you, if
16:04
you take this case approach, you say, okay, I'm going to break my
16:08
test data into easy medium hard basic thing.
16:12
And then cause, you know, I don't know the problem yet.
16:14
And then you go easy, medium hard.
16:16
And then you push the production, you see problems
16:18
come back and then you say, you know what, I need to look at
16:21
different kind of evaluation. So I want to see different
16:24
orientation of the car, different to crucial level
16:27
camera conditions and weather conditions.
16:28
These are basically trying to do like 10 to 15 test cases and
16:32
then you learn more about you. You go ahead and the more you
16:34
working on the improving the product and then you come back,
16:37
you see, bugs, you start breaking these down into more
16:40
test cases, more fine-grained more.
16:42
Resolution test cases. Now, that's the year end.
16:45
You start with 3 ends of the 1020 hundreds of those and then
16:50
an engineer joins in a year. Now, they have been working on
16:54
cars vectors. So you understand exactly what
16:56
we're looking for. We, we always as, you know, data
16:59
scientist, we look at the end when you evaluate the model or
17:01
on, when you change your look at the data, one by one, don't have
17:04
a sense of what this is like, you look at a car.
17:07
You're not, you don't know what you're looking at, you know,
17:08
it's a car. You don't know.
17:10
It's like you care about recruiting you care about
17:12
Weather conditions are so on and then the deployment environment
17:16
is different for each customer. If you of, it's like a robotics
17:20
kind of application indoors, you don't need some kind of weather
17:23
conditions, for example, others, you need up.
17:25
So to understand the product specifications understand that
17:29
there are deployment environment.
17:30
The customer environment, understand what the team has
17:32
been working on it. All happens in that knowledge
17:36
base knowledge base starts from just looking at the list of test
17:39
cases and looking at the history of how things happened.
17:43
Being able to, I didn't understand within the first
17:45
couple weeks. This is what these are the
17:48
customers that I am building for, and these are the specs
17:51
that we care about So the one thing that comes to mind when
17:57
you're talking about like tens and hundreds test cases and
18:00
things like that and I can see that y z value we need but do
18:04
you or have you experienced any issues with actually the
18:07
computational cost of running? So many tests and things like
18:10
that? And like building all the cost
18:12
of building data sets for the purposes of testing and things
18:15
like that because I can imagine that you're adding loads of
18:17
robustness and scalability, right?
18:19
But what significant computational cost I guess.
18:23
No. Absolutely.
18:24
And actually, we are pushing for more testing.
18:27
I were pushing for much much larger test data which could be
18:30
intuitively increased the the computer production cost, and
18:35
storage and, and effort into, on the way we solve this problem is
18:40
this Avast or I know you guys talked several times in your
18:43
podcast about the dollar store and we are fans of their of the
18:46
concept. When you're testing on, let's
18:49
say, hundreds of thousands of images and you run the model and
18:52
then 50,000 more images. And then the positive thing
18:58
about having a platform to do that is you do you need to run
19:02
less emphasis because the model has been tested on the other
19:05
image before. So you look at the nucleus item
19:09
in this in this process is the image or an object with its.
19:13
Would you frame or something? You look at the image on your
19:16
create, we create a history of this image.
19:18
So at the end of the day, when you look at an image, you see
19:20
that this image has been tested on these 10.
19:23
Models. And this has been, the history
19:24
has been promising. So so now the when you're
19:27
testing and you re testing then if this version of the model has
19:30
been tested on this version of the image or the data sample, if
19:33
it's different than if it's infinity or other side, then you
19:37
test less. So you are testing more but you
19:40
are running a much, much less computation and comes here.
19:44
They're reporting these most of the testing happens not because
19:48
not because the engineers are validating their product, it's
19:51
answering questions and being able to Compare models that
19:54
happens once. So we saw is actually very
19:58
practical team spends. Like if you're running emphasis
20:01
on millions of images, you get like spent a couple of hours on
20:04
that and then it's up there and then every time you run, you add
20:08
more data. You run anything on the model.
20:10
It just runs the addition they do this.
20:13
So we saw the 5x of the 10x. Even love.
20:17
Our current customers is 10x that the size of the data and
20:21
the number of test runs that they're running with much lower
20:24
competition costs, then before, And so you talk because it goes
20:30
back to the exact question, Adam.
20:32
You were asking of whether it's a tool on a platform because
20:37
testing is not liked, and it's always an afterthought and to
20:40
tell people you guys, who just think we found that it has to
20:43
be, you have to solve the whole thing.
20:44
We cannot solve one piece at a time.
20:48
Let's riff off of that really fast because I think it's a
20:52
theme and everyone knows it. The testing is not liked and
20:56
nobody wants to hear that they need to do more tests.
20:59
Do you have silver bullets for someone like let's say, Adam
21:03
goes tomorrow and he says, listen everybody on his team, we
21:08
got a test. Like, we're I just had this
21:10
podcast yesterday with Moe and I realized that we're not testing
21:14
enough. What do you do to get by in how
21:17
do you, convince people to go along this journey with you?
21:22
Yeah. We've been, I've been doing this
21:24
myself in my career and then I'm doing it every day today.
21:28
Alright thank you for giving your number one is the product
21:31
we see now that most of quality initiatives start from person in
21:38
the team, technical leader a manager.
21:40
Some that is pushing this quality levels.
21:43
That's number one, or you have been getting a lot of pain from
21:47
production, from sorry, from customers.
21:50
And then customers are pressing on business.
21:52
This is passing on engineering, so we we try to look at the
21:56
team. We, this is, are the two main
21:58
things we focus on. I tell Engineers or they are
22:01
scientists. Hey, you, but you have have a
22:04
slide on this specifically. This is first slide.
22:06
We talk about all the time is if you test more or just right, if
22:11
you have a systematic test process you will train this.
22:15
You will do less data collection and data labeling.
22:18
You will have to answer a lot less questions.
22:22
It's from people because the answers already rather there are
22:26
scientists usually like the are skeptical or Engineers Capital
22:29
until the see the end to end of how we built it and they start
22:33
seeing okay we see your point and then we do it like a pilot
22:37
or something the see the numbers before like how many times he
22:39
tested let's say before was once a week.
22:41
Now, 10 times a week but the actual hours and testing are
22:44
much less because they just like it, did they run a lot of tests?
22:47
And the actual hours are spent on the training which is very
22:51
painful process. Shooting in the dark kind of
22:54
process. All this got gets myself, my
22:57
pitch to data scientist is like a, you don't have to spend all
23:01
this time. You'll spend less time on
23:03
testing as well. And you will have a systematic
23:06
process that everybody will agree on a product as well for
23:11
our teams and business teams so I can certainly get that right?
23:16
Yeah. Can see the actually the point
23:18
about being a platform and how you then save or more runs And
23:22
you save time, holistically makes a lot of sense.
23:25
Actually, one thing that, so the next thing because I'm my brains
23:28
working, right? But my, the next thing that
23:30
comes to mind in is the so someone like me that came in
23:33
today, science from a physics background getting started with
23:37
testing. I've got these software
23:38
Engineers, are my ears, like young data, scientist, getting
23:41
started. It can seem quite alien and
23:43
quite like I'm really get it EDD.
23:45
How do I do that? And so, And actually, as you go
23:50
further in your career, many people never get over that
23:51
hurdle of, they just don't get it.
23:54
And I've seen so many teams where they just, we don't do any
23:56
testing a fight. Okay.
23:58
So, because I'm aware of it, like you're testing, is only as
24:02
good as your quad. You've asked a bit of a skill
24:04
like testing. He's a bit of a skill.
24:06
So do you have any recommended resources or like routes for
24:09
people to get confident with this, or get good at testing?
24:13
And like how should they be, what approaches should we be?
24:16
Taking like does tdd work and stuff like that.
24:19
I'd love to hear your thoughts. Absolutely, starting notice
24:22
which we try to avoid the word TD having been very
24:26
controversial. It is the he do.
24:28
I mean, it is what it is. It's just, we called it
24:31
eccentric development. And the reason tdd is is a
24:36
little bit controversial in the software side because it's more
24:39
of a deterministic. It's not a statistical process
24:42
is basically compared to ml. It's really hard to infer.
24:48
It's based on data sets or benchmarks.
24:51
If you're looking at a human level performance, to give the
24:54
human set of data, and they are accurate in 99% of the data, you
24:59
can infer that. This one percent is the most
25:00
adverse are case for a human level, but you cannot do the
25:04
same thing for a data product. Like ml, you have to actually
25:10
include every piece of it. Also, I forgot to question.
25:16
Yeah. So that now that's the
25:17
interesting in itself, the tdd kind of, because I've heard
25:20
loads of mixed views on that. So the next part for me was,
25:24
yes, but I don't want to get good at testing and how do I as,
25:27
like, from a machine learning point of view, what things?
25:31
Yeah. So for that we testing, is
25:34
basically done now by every data scientist.
25:36
Our actual that the mechanics of it, which is running emphasis
25:39
and getting results. So that is not changed and What
25:43
we are changing is the way you're managing your data, your
25:46
test data, to things, test data has to be has to cover the
25:51
entire span of what you're trying to achieve sounds
25:54
impossible. But it's actually much much more
25:56
manageable than trying to do that on the training data test
26:00
data. If you're, if you're trying to
26:02
build a product from steps, you have a say, I wanted to be able
26:06
to take this class with these kind of sex.
26:09
You go backwards and you create that play with your data, create
26:12
these the test cases. Now, these test cases are the
26:16
way they are crazy today, whether they are images or text,
26:18
there are quite a lot of data wrangling.
26:21
We built the studio for that. It's a UI that you're able to
26:24
stratify your data search your data, based on any snotty that
26:28
you're looking for without prior labeling.
26:31
So this is our whole UI thing and then you're able to add some
26:35
augmentations to test robustness and include some objects.
26:38
So we created like an interesting intuitive.
26:43
Video that we looked a lot about the u.s. piece of it.
26:47
How to make these Engineers of their scientists coming from
26:51
managing background. But the understand data, they
26:53
understand how what, what to look for in the really
26:56
understand what neural networks are.
26:58
And then we start building UI on top of this.
27:01
And then we also dealt the an SDK to make it Partners UK to
27:06
make it so easy for them to integrate similar to what some
27:09
biases. So to integrate into their
27:10
current workflow. So that you experience is Is the
27:13
biggest piece into making sure that this there's adoption into
27:17
this new workflow, because we're building a bit of the day,
27:20
building animal testing tool. It's what starting a new
27:23
category of two Wings. We're told, we're not telling
27:25
people how you've done and we're actually we found that we're
27:28
changing the behavior and I believe that the workflow will
27:32
change after. If that does Centric approach is
27:36
created the way you're managing your test data will reflect into
27:39
your whether you're managing your training data, your flex
27:41
and everything up straight. Their scientists are already
27:46
familiar with the pieces that we were working with and we cut off
27:49
all the data wrangling approach that they have to reload process
27:53
and understand the and even like the manual approach of having to
27:57
look at the data one by one labeling.
27:59
This approach, we cut all these hurdles because we not, we think
28:03
a lot when we're working on testing with the about the
28:06
statement that we started with the way wants to be told to test
28:09
a lot, right? So we try to make sure that this
28:11
is easy and where everything was Measuring.
28:13
We're testing, how long it's taking you and how easy it is
28:16
compared to current approaches. And so there's one side effect
28:23
that I think we haven't touched on but it kind of gets me
28:27
excited about the way that you're looking at the problem.
28:30
And that is considering right now, both Adam and I are in the
28:34
EU and there are proposed regulations that are coming out
28:40
your machine learning and AI, is they Like to call it in the EU.
28:45
And so, I feel like this, have you thought about like how this
28:49
and how the idea of testing /, creating more of a knowledge Hub
28:55
and better, documentation can play into the regulation aspect.
29:00
Yeah. Actually just would have thought
29:04
we have thought about it. We Disturb mode that we are at.
29:06
Now at Kelowna, we are working with specific teams that are
29:11
facing regulations and independent regulators and
29:15
independent Auditors into the process.
29:18
So we start with, if you look at any in domain now, it's getting
29:22
clear regulated and then they are telling you, you'll see if
29:25
you're working on a face album. First section problem, you have
29:29
to show gender bias Race by Sage advice.
29:32
These kind of metrics that you can just say like 90-something
29:36
percent in detecting faces. So and let the same thing for
29:40
safety applications and more mobile Mobility applications in
29:43
TV and so on. Now this all starts from the way
29:47
the regulators and we want heavily with this Regulators
29:49
there. The way they and their great
29:51
example of this is Nest, the gulf.
29:53
If you guys are familiar with the it's a u.s. nonprofit
29:56
organization. So basically they are the way.
29:59
The, if you look at the process of Open Source, you look at the
30:01
reports they care. A lot about these, what we call
30:04
Stacks again. I look at regulation
30:06
requirements as requirements aspects and then you look at
30:09
these regulations D1 test biometric differences or any
30:14
other, luckily they care about and then you start implementing
30:16
that in your specs. This secret test cases, that's
30:19
your model identified failure modes and then trained against
30:23
only these specs. So we're trying to eliminate the
30:26
shooting in the dark kind of process in the training.
30:28
So it starts from there like these three of relations are
30:31
requirements. Isn't he?
30:33
After that of you're done this, it's so easy to create specific
30:36
reports for them basil, you have already there the results.
30:39
You create specific reports and custom Force for these
30:42
organizations. It's a common problem that their
30:47
regulations thing that gets we started with actually areas that
30:50
are heavily regulated because the need hear their story.
30:54
They are looking at testing to be able to answer these
30:56
questions. There's that pain that you were
30:59
talking about earlier and how you see ya the different it
31:03
being reflected in the business. So I want to change gears a
31:08
little bit. Well I mean it's not completely
31:10
off topic but one thing that testing does and we spoke about
31:15
before Is identifying failure modes, and yes.
31:19
So I wanted to dive a little bit deeper into that.
31:22
And how does testing identify these failure modes and get this
31:28
granular understanding of the Model Behavior for each
31:32
scenario? I'm sure this is a that's the
31:36
core of what everything would promising is.
31:39
At this point, you want to share some slides.
31:41
I put couple slides to answer like to give a background of
31:44
what this what failure modes are?
31:47
Yeah, throw it on for sure. Yeah, this is actually they're
31:50
like we start with this and that's a failure mode.
31:52
Like, well, it's from our previous example, like I should
31:56
we have, we were building a compact gun detector and we used
31:59
the luckiest and months improving the statistical
32:01
evaluation metrics that we have and then we can Up with like a
32:04
model of 99% accurate, on the corner of F1 score and then we
32:10
did not find like we put it in the demo and we that gun was
32:13
that clear? And then we missed it in a very
32:15
easy case. And that is a failure mode that
32:17
was hidden but did not understand that this was a film.
32:21
Would you just assume that 91% that's missing, is probably the
32:24
rest of the most, a staircase and we didn't notice that.
32:29
And then the problem here happens again, goes to the
32:32
failure modes, you were not able to identify Five failure modes
32:34
long as we're using the current evaluation processes, which
32:38
basically if you look at it today, we have some Benchmark
32:42
data sets. We keep making it larger include
32:44
increasing into it. We run emphasis, and we spend
32:46
days or even weeks in some applications trying to
32:51
surrogate. Errors are gaps and evaluating
32:53
the model. So yeah, we're currently at the
32:55
current evaluation process. We Run emphasis on large,
33:01
arbitrary created, Benchmark sets and then we spend days or
33:05
even weeks staring at their charts, try to understand the
33:08
model performance. And then that involves a lot of
33:11
images like floral or data samples to look at the failures.
33:15
If you have a million images in your test set and you feel 1%
33:18
that's 10,000 images, right? So you can understand everything
33:22
and write down notes and so on create reports and then you at
33:24
the end you you just produce some magical Global metric,
33:28
right. Then doesn't tell you exactly
33:30
what's happening. So this is the world that we
33:32
live in today and biscuits value in teams that want to evolve in
33:36
the quality side. These just document this Summit
33:40
somewhere and then try to manage that somewhere and then you see
33:43
here, two main problems. These Frozen benchmarks and the
33:47
global Matrix and this takes us to that problem.
33:50
Beheading circulation problem, which goes directly to your
33:53
point about failure modes. So these classes, this like you,
33:58
Global metrics for the class. They don't like we said, we
34:02
don't capture the full picture and then the hidden
34:04
certification problem. Just basically says that these
34:06
models, they exhibit highly valuable performance for some
34:10
different subclasses within the class.
34:12
And then the office of the of the Hidden stratification.
34:14
People just share the other example from a different domain
34:17
which is x-ray. Just x-ray images.
34:20
They have been measuring this blue curve is the curve, which
34:23
is the aggregate or the global metric, only to find out that
34:26
they slice. The curb out of a slice of they
34:31
care about is that green slices, which is noches doesn't vote
34:36
suggestions. So that was failing silently and
34:39
the same gun examples and this applies to every domain and demo
34:43
site. So to your question, limiter is
34:45
the problem in identifying failure modes happens and the
34:48
way you are managing your data, we want to take a test case
34:54
approach in managing the data. So I'm going to skip this but
34:57
this is basically Lee what we're pushing for.
34:59
We are pushing for a paradigm shift, from a world that we live
35:04
in today or that involves a lot of large benchmarks arbitraries,
35:08
like data into more of a fine-grained well-curated test
35:13
cases. So in the gun example here,
35:14
instead of throwing a lot of gum examples there, just stratify
35:17
those in the six or seven of the start.
35:20
You got here by say, hey, I'm building this product.
35:22
I want to be able to identify a gun in empty bin like what we
35:24
had and different views and you add to that.
35:28
Front camera conditions and sensors because that affects
35:31
family there in your natural performance, and then you have
35:35
the what we like to say that high resolution test results and
35:38
then you're able to compare models and in different runs
35:41
based on these criteria that you care about.
35:44
So if we are from looking at this Global metric, we had a
35:48
model that had ninety six point. Eight percent that improve 299.
35:52
We deployed the 1990s because that's all we know what.
35:55
If you look here, we found out that we pushed We made a push
35:59
when we are going from A to B to improve on the occlusion side
36:02
and we have regressed an easy case out of the area.
36:06
So that is what I meant by fit. Failure mode, the funding
36:09
pass/fail criteria. We thought it's a hard thing to
36:11
do. It's not that hard.
36:13
It can be done and that is the only gate.
36:16
So a fully automated workflow. We never deployed without.
36:20
Yeah. Right for just a heads up for
36:24
the people that are only listening.
36:25
We've got two different columns here.
36:28
And you see Model A which is actually a lower accuracy but it
36:32
ticks the boxes more than model B which has this 99% accuracy
36:38
but it's actually failing quite significantly on more of these
36:43
boxes like the gun in the empty bin or in the the different
36:50
except of you that you're testing for, right?
36:53
Yeah, absolutely. So.
36:56
And then with this, if you look at that approach, let's talk
36:59
again about why this makes everything above and Upstream
37:02
process less or it is training less.
37:05
So on now as at the simplest form of it, if you have a fleet
37:08
of models which we always have to have tons of models that
37:11
we're sending and then don't just have them sit on their AWS
37:14
bucket or gcp bucket. Now with the just run emphasis
37:18
and store them on: and now you have a new scenario you have a
37:22
new customer the same this and be clear about their environment
37:25
has Types of sensors and these types of scenarios, then you run
37:29
all your models on this scenario and then find the best to
37:33
deploy. So you in many cases we've seen
37:35
that, you don't even need to train a new model.
37:37
The idea of having one General model that is great at
37:40
everything, it's really painful. And then you find yourself
37:43
playing like a whack-a-mole like you crush a bug and then you go
37:46
ahead and pop comes up. So what we ended up doing is we
37:50
said okay this model is great for demo.
37:52
Demo is a controlled environment.
37:54
Let's make let's give customers. Great.
37:56
Deals first and now another mother's transferring was okay.
37:59
This types of extreme issues of this model of this kind of
38:03
sensors have this event. You don't have to build a model
38:05
that is great for all sensors or great.
38:08
For all scenarios. Have a few months and then on
38:10
top of that, you can even end up being having an example of
38:13
models, of our greater different things.
38:14
So, the mindset of building Animal product completely
38:19
changes from the mindset of what we have today is that lust, with
38:22
more data. Let's make this model brilliant
38:25
that every And then I want to show like this is nice thing
38:30
here. Is that is what the workflow
38:31
looks like we're looking at. This is what you call the test
38:35
Centric, development or production, ml.
38:38
We start with the, with this product specs, it has to be,
38:41
what are you building for? Let's write the product specs
38:44
and create these test cases. And then when you click this
38:47
cases, test them, and then you start here, improving on it,
38:51
like the symbol training data, and train, and test assembled,
38:54
translate and train, and test. So it's all All like, this is
38:57
the Target that you need to find the target before you take a
38:59
swing, right? So we set the Target right here,
39:01
and you keep taking swings until you're ready.
39:03
And then when you observe a bug in production, go back to your
39:07
test, create the test cases, like this mug again like the
39:10
gun, somebody that we saw create a test case, that's your models.
39:13
And then assemble because the solution, we what we end up
39:16
doing today, is we have a problem, let's create more
39:20
training data, more training data, and then to test.
39:23
Whether the model is good at is not.
39:24
You need tens of probably. Just like to train a model in
39:28
snow, you need much much larger data, so and then the solution
39:31
is not always, let's get more data.
39:33
Could be tuning, could be augmentation could be different
39:35
things. So training the the collecting
39:38
data and labeling is a very expensive piece, very painful
39:43
and it's always solved at the labeling peace and automating
39:46
the level in peace. I believe that it can be heavily
39:49
reduced by working backwards from the goal and going through
39:53
the door with their life cycles. So for example, interesting and
39:57
again I'll just for people that just listening to this.
40:00
So presented here is like an animal Productions or
40:02
development workflow. The steps through getting your
40:06
business of Jason respects through to then create some
40:08
tests and and then yet testing of almonds and shipping in so on
40:12
and so on. But it has too much feedback
40:16
modes. One is the one you, you
40:17
typically see in these kinds of diagrams of, get your training
40:20
data, trying to models, test the performance and feed that back
40:23
into you here sort of threshold. And then but the other loop is
40:28
probably unique to this diagram from my perspective.
40:30
Anyways, the sort of once you've deployed a model to production,
40:33
perhaps you look out for Boggs to identify and then go back to
40:37
create tests to catch them and and so on and so forth, which I
40:40
think the really interesting way to do this.
40:43
But that one thing I wanted to ask and it kind of inspired by
40:46
the the diagram with the test case is was this.
40:50
I can see this working its kind of, after the fact a little bit
40:53
though, is there a way to identify Defy scenario.
40:58
I think I know the answer to this but I guess the only way to
41:00
identify scenarios he's through. Some really vigorous
41:04
brainstorming like up from or through experience and domain
41:08
knowledge and expertise it, there's not another way doing
41:11
that that's quite natural is that that's exactly what it is.
41:14
And this is where the knowledge base comes in handy.
41:17
You have to go through this, it's either personal experience,
41:20
or like, for the product manager or they're the team that is
41:23
unseen. Paul comes from iterating Over
41:25
the product. Either way, after some time, you
41:28
are learning more about your domain.
41:29
Even if you are working the same domain, but different types of
41:32
customers is completely, like, yo, you don't understand exactly
41:35
what you're testing for. So we believe that everything
41:37
starts here either. The product specs for the test
41:40
cases, test cases of, when I have a new in new data
41:43
scientist, joining the team. First thing is like look at the
41:46
test cases that we have and look at the history of models.
41:48
How we improve that we will always been good at these
41:51
scenarios. We don't we can us, but we
41:53
always been bad. And then you start managing, you
41:55
know? The whole biggest problem I hear
41:58
about ml is, how do you product manage it?
42:01
How do you always ask them all team?
42:03
And they tell you, it's an iterative process maybe in some
42:06
time which the right food. So like this, it's a painful
42:09
process, but we want to make sure that the cell is like a
42:12
software engineering product. It can be product managed.
42:15
You can say, we see our customers saying that, okay, in
42:18
there are okay, are the next quarter will go where you need
42:21
to improve on these five scenarios.
42:23
Make sure I am not regretting what we have.
42:26
When you're actually incrementally systematically
42:29
improving. And every time is actually
42:31
leading to real Improvement towards the target, which is the
42:36
product or the customer I really love that is over because it
42:40
again puts the end use of right at the very core of what, people
42:44
perceive it to be a very techie process.
42:45
But actually like that feedback from the expert, the end user
42:48
insights becomes even more important.
42:51
In addition to puts the customer the end user, which the product
42:56
and business team heavily involved in the process, this is
43:00
where we have to look to the. The point is, if it can't be
43:03
just that Dev tool. We, if you want to build animal
43:05
testing, we believe that everything starts from that from
43:08
testing. So after building a UI for and
43:11
some executive dashboard and manager dashboard to manage,
43:14
even the team's performance of this is like a version of
43:18
Deathstroke last year, I kind of think to be able to manage these
43:21
test cases and all the kind of stuff that we jump through
43:24
tools, as data scientist all the time and to go to teens and
43:28
children change what you're doing.
43:30
We have to make sure that southern hate everything in.
43:33
There's as much easier than having to jump through different
43:35
tools. It's so cool.
43:38
I love seeing this. And so, there is one thing, I
43:41
think we should probably highlight to.
43:44
You've been talking a lot, about computer vision examples.
43:48
How does this differ when it's on structured data?
43:53
And of data like what, what are the big things that you see?
43:59
It's easier just to turn my skin off.
44:03
So it gets a little bit easier and in Text data table data, not
44:10
obvious. But the problem the biggest
44:12
hardest is why this what we're talking about is not
44:15
implemented. We talked to hundreds of people
44:17
of teams especially last year and for that the biggest pieces
44:21
like how do you Mid this, paradigm shift from these data
44:25
sets into scenarios because you, you can collect for these
44:29
sliders. But usually, this is not the
44:31
starting point, as you starting today.
44:33
When we have millions of images in the test set, how do I
44:35
stratify that? Now, what our solution for this,
44:38
we not a image retrieval system or we like to look at it as More
44:43
Of A recommender Plus image retriever system that you'll be
44:46
able to search for scenarios. Like if the three of us are
44:50
looking at our women were here, the recession chemistry.
44:52
Data. Look at an adult with a beard.
44:55
Probably the three of us will come up and then I'm going to be
44:57
adding headphones. Only, I will come up, right?
45:00
So basically, you're able now to search your data, for for the
45:04
exact scenario is you're looking for and then there's a lot other
45:08
piece of in that studio test case study, we had a whole piece
45:12
for 40% of our product, is that this this studio, it's a secret
45:15
as a creative space. You will go in there and then
45:18
play with the images stratify augment at, you know.
45:22
It kind of more scenarios, I'm doing create a test case for the
45:26
minutes. Now that is much easier in text
45:29
searching, text is much easier. Augmented text is much easier so
45:33
creating these and even lower computational cost as well.
45:37
So we started with vision because we believe that is the
45:40
hardest problem. We are spent most workers in
45:42
this area. We actually are building our NLP
45:45
product will be out by March and it's we seeing as it's the test
45:51
case. ER is the same being able to
45:54
manage and version test cases and Pad mode and edit and so on.
45:57
But being able to the studio is much easier than going with
46:01
images. So the structured data like
46:03
tables and Text data is easier to go through stratifying peace.
46:11
Sweet. I can't let you leave here
46:14
without you talking about your book because I think that's how
46:18
I first started like following you and I found out about you
46:23
can we go into that a little bit and just tell us like so the
46:26
book is for anybody, that doesn't know.
46:28
It's machine learning for vision systems.
46:30
What what can we expect there besides?
46:34
What is on the tin? Yeah.
46:37
Yeah. So the book is called Deep
46:38
learning for vision systems. And it starts after you have
46:42
understood what the machine learning goals for workflow,
46:44
looks like and the story of this happened like again, few steps
46:49
before I was at Amazon and I was at the Kindle organization.
46:53
I wanted to work on the ml side. I found a gap in the CV Gap and
47:00
inside there was back 2014 2015. I found a gap here in be and
47:04
having dark a CV Authority in the organization.
47:08
So I built I found the car and organization CV think tank for
47:13
four different teams and Amazon to start with that to start
47:17
building a team. It's, you know, how hard it is
47:19
to higher energies was much, much harder back then.
47:21
So my Approach was. Okay, Amazon has a organization
47:25
called mlu machine Learning University and it's for Amazon
47:29
Engineers. So, I kicked off a course and
47:32
the Seedy side and I we had like 30 engineers and room and then
47:36
we started with 55 of those like, after a couple months and
47:39
read that was Heat of the team. So the Project's initially
47:42
started as that and then from creating the a lot of material
47:46
for and I work with the dusty and creating a monell degree and
47:51
and telling different organizations.
47:53
So, from creating a lot of content, I realized that the
47:56
have it's all told in from one angle which is same workload.
48:03
They're trained test and deploy, but there's a lot of and then
48:07
some other heavy on math summer are heavy on applications and my
48:12
belief is math. My belief is like the you don't
48:16
need. You need a little bit of math
48:17
just to understand what's going on.
48:19
So I don't like you. I don't want to admit it but you
48:22
don't need to take a calculus course or some kind, of course,
48:25
it's always starts with my pack plus 101 or linear algebra, 1 0
48:29
1. So, I inserted this math, it's
48:32
very practical book and I wanted people to understand this map
48:36
pieces for one goal to be able to It and Implement research
48:41
papers. So all books are going through
48:44
explaining Concepts. I was trying to train the
48:48
readers to learn on their own. I have a full chapter after I
48:51
explained these fundamentals I have a full chapter just going
48:55
through maybe like seven or eight papers one by one, their
48:59
second them and creating the coach for them just to after
49:02
that because you understand that ml just evolving heavily the
49:06
books, you get obsolete fucking a couple years and the goal
49:09
here. It is not really to share
49:10
information. It's more to teach people each
49:14
Engineers, how to go ahead and learn on their own read papers
49:17
and stay up to date with the confirmation.
49:19
So hopefully to answer your question Demetrius.
49:21
Hopefully after reading the book you'll be or others be able to
49:27
go ahead and read any paper as much as they can and then being
49:32
able to implement it understand what they're talking about and
49:34
implementing code. So cool.
49:37
So we'll leave a link to the book.
49:40
If anyone wants to get their hands on it and mow, this has
49:43
been amazing. I love talking to you about this
49:46
new paradigm. The new idea of him.
49:50
As you do not like to call it tdd.
49:53
That is a bad word in your it's a four-letter word, I guess.
49:59
Yes, test Centric, development. It's so cool to hear about all
50:04
these different ways that people are.
50:06
Talking the problems and how we're trying to evolve the space
50:10
that we're in right now. It's obvious, that there is a
50:13
lot of pain in this space and it's obvious that a lot of
50:17
matured maturing needs to happen.
50:19
So that's why I wanted to get you on here and talk to you
50:22
about how you see this, maturing and evolution happening over the
50:28
next two to five years and it's been great.
50:32
You definitely tick the Box on that one ice I appreciate it and
50:37
Adam, thank you for coming man. As always it's a pleasure.
50:40
Having you here co-hosting with you?
50:42
No, thanks very much and good night.
50:44
I've kind of just read a lot of stuff about you're deploying a
50:46
know, engineering guidelines, and I have to rip.
50:49
So that up, I think and revisit some of this testing stuff.
50:53
Oh, that's how, you know, it's good man.
50:55
That's how you know, it's real good.
50:57
So you made you think about it again.
50:59
Sweet Muhammad. I appreciate it that we're gonna
51:02
end it here. That's good conversation.