Up next in 10
How to Prevent AI Model Drift: Continuous Retraining for Image Classification Systems
Aug 6, 2025
AI models, especially those used for image classification, face performance degradation—also known as model drift—when deployed in real-world environments. Accuracy can decline by up to 15% within just three months, leading to major operational and economic setbacks.
In this session, we explore how modern convolutional neural networks (CNNs)—from AlexNet to EfficientNetV2—are affected by model drift, and what strategies can prevent it. You'll learn how leading organizations are implementing continuous retraining pipelines, using selective data sampling, adaptive triggers, and validation protocols to:
• Extend model life by 42%
• Reduce retraining costs by over 60%
• Maintain over 92% of peak performance with only partial retraining
• Cut false positive rates by 43%
Whether you're working in healthcare, manufacturing, or any AI-powered domain, this talk will provide practical, scalable retraining strategies to ensure your models stay accurate, reliable, and production-ready.
⚡Download Sharp Rewards Wallet to earn Rewards: (200 bonus on using this link) http://invite.sharpplatform.com/VirtualConf
📺 CSharp TV - Dev Streaming Destination http://csharp.tv
🌎 C# Corner - Community of Software and Data Developers https://www.c-sharpcorner.com
Show More Show Less View Video Transcript
0:00
Hi everyone and welcome to a new episode
0:01
Hi everyone and welcome to a new episode
0:02
Hi everyone and welcome to a new episode of AI Dev Tools. Today I'm your host
0:03
of AI Dev Tools. Today I'm your host
0:04
of AI Dev Tools. Today I'm your host Simon and we are back once again. In
0:06
Simon and we are back once again. In
0:06
Simon and we are back once again. In this episode we are going to learn about
0:08
this episode we are going to learn about
0:08
this episode we are going to learn about preventing AI model drift continuous
0:10
preventing AI model drift continuous
0:10
preventing AI model drift continuous retraining strategies for production.
0:12
retraining strategies for production.
0:12
retraining strategies for production. And for this episode we have Amit Aurora
0:15
And for this episode we have Amit Aurora
0:15
And for this episode we have Amit Aurora who is a software engineering manager.
0:21
[Music]
0:23
[Music]
0:23
[Music] Hi Amit and welcome to the show.
0:26
Hi Amit and welcome to the show.
0:26
Hi Amit and welcome to the show. Hello everyone. Um and thanks for
0:29
Hello everyone. Um and thanks for
0:29
Hello everyone. Um and thanks for watching and joining this one. Um as
0:31
watching and joining this one. Um as
0:31
watching and joining this one. Um as Simon said, I'm Ahmed Dora, software
0:33
Simon said, I'm Ahmed Dora, software
0:33
Simon said, I'm Ahmed Dora, software engineering manager. So I'll be talking
0:35
engineering manager. So I'll be talking
0:35
engineering manager. So I'll be talking about the topic that's increasingly
0:37
about the topic that's increasingly
0:37
about the topic that's increasingly important as the AI system move from the
0:39
important as the AI system move from the
0:39
important as the AI system move from the research labs to real world deployment.
0:42
research labs to real world deployment.
0:42
research labs to real world deployment. Uh how to prevent model drift and image
0:44
Uh how to prevent model drift and image
0:44
Uh how to prevent model drift and image classifications using continuous
0:46
classifications using continuous
0:46
classifications using continuous retraining strategies. Uh the
0:49
retraining strategies. Uh the
0:49
retraining strategies. Uh the presentation is based on my recently
0:51
presentation is based on my recently
0:51
presentation is based on my recently published article in the journal of
0:52
published article in the journal of
0:52
published article in the journal of computer science and technology studies.
0:54
computer science and technology studies.
0:54
computer science and technology studies. I'll try to share real world evidence uh
0:57
I'll try to share real world evidence uh
0:57
I'll try to share real world evidence uh practical architectures and
0:59
practical architectures and
0:59
practical architectures and recommendations to keep model performant
1:00
recommendations to keep model performant
1:00
recommendations to keep model performant and trustworthy in production
1:01
and trustworthy in production
1:02
and trustworthy in production environment. We're going to talk about
1:03
environment. We're going to talk about
1:03
environment. We're going to talk about how do we prevent model drift in image
1:06
how do we prevent model drift in image
1:06
how do we prevent model drift in image classification using continuous
1:08
classification using continuous
1:08
classification using continuous retraining strategies. Uh I'll move to
1:10
retraining strategies. Uh I'll move to
1:10
retraining strategies. Uh I'll move to the actual content. Uh first talking
1:14
the actual content. Uh first talking
1:14
the actual content. Uh first talking about recap of how far we have come as
1:17
about recap of how far we have come as
1:17
about recap of how far we have come as far as the
1:19
far as the
1:19
far as the accuracy of image classification is
1:20
accuracy of image classification is
1:20
accuracy of image classification is concerned. I think in 2011 the top tier
1:24
concerned. I think in 2011 the top tier
1:24
concerned. I think in 2011 the top tier accuracy on standard benchmark was
1:27
accuracy on standard benchmark was
1:27
accuracy on standard benchmark was around 50%.
1:28
around 50%.
1:28
around 50%. Uh and then alexnet brought it to 63% in
1:32
Uh and then alexnet brought it to 63% in
1:32
Uh and then alexnet brought it to 63% in 2012 and by 2023 architecture like effet
1:36
2012 and by 2023 architecture like effet
1:36
2012 and by 2023 architecture like effet v2 uh pushed it over 90%.
1:40
v2 uh pushed it over 90%.
1:40
v2 uh pushed it over 90%. uh despite these advances, production
1:42
uh despite these advances, production
1:42
uh despite these advances, production model deployed in real environment often
1:45
model deployed in real environment often
1:45
model deployed in real environment often see performance decay due to evolving
1:47
see performance decay due to evolving
1:47
see performance decay due to evolving inputs and context and that's where the
1:51
inputs and context and that's where the
1:51
inputs and context and that's where the uh drift becomes critical.
1:53
uh drift becomes critical.
1:53
uh drift becomes critical. Moving on to the next uh
1:57
Moving on to the next uh
1:57
Moving on to the next uh slide the I I think without like if you
2:01
slide the I I think without like if you
2:01
slide the I I think without like if you don't do the proactive maintenance model
2:03
don't do the proactive maintenance model
2:03
don't do the proactive maintenance model accuracy roads by 1.5 to 3% every month
2:08
accuracy roads by 1.5 to 3% every month
2:08
accuracy roads by 1.5 to 3% every month and can collapse by 15% within a
2:10
and can collapse by 15% within a
2:10
and can collapse by 15% within a quarter. Uh enterprises are now
2:13
quarter. Uh enterprises are now
2:13
quarter. Uh enterprises are now allocating around 30% 35% of their ML
2:16
allocating around 30% 35% of their ML
2:16
allocating around 30% 35% of their ML budgets just to model maintenance. So
2:19
budgets just to model maintenance. So
2:19
budgets just to model maintenance. So the takeaway here is very clear. Um so
2:23
the takeaway here is very clear. Um so
2:23
the takeaway here is very clear. Um so deploying a high accuracy model isn't
2:26
deploying a high accuracy model isn't
2:26
deploying a high accuracy model isn't the finish.
2:28
the finish.
2:28
the finish. Basically it's it's the starting point.
2:30
Basically it's it's the starting point.
2:30
Basically it's it's the starting point. That's where you start when you once you
2:32
That's where you start when you once you
2:32
That's where you start when you once you have deployed your uh high accuracy
2:35
have deployed your uh high accuracy
2:35
have deployed your uh high accuracy models.
2:40
Um so this chart this is very
2:43
Um so this chart this is very
2:43
Um so this chart this is very interesting chart. It shows how the
2:46
interesting chart. It shows how the
2:46
interesting chart. It shows how the static models degrade over time. Like
2:50
static models degrade over time. Like
2:50
static models degrade over time. Like you can see the false positive increased
2:53
you can see the false positive increased
2:53
you can see the false positive increased by over 7% per month and false negatives
2:57
by over 7% per month and false negatives
2:57
by over 7% per month and false negatives are up by like almost similar like 6.8
3:01
are up by like almost similar like 6.8
3:01
are up by like almost similar like 6.8 percentage every month. So I mean and
3:04
percentage every month. So I mean and
3:04
percentage every month. So I mean and these just aren't the numbers that
3:06
these just aren't the numbers that
3:06
these just aren't the numbers that models are degrading right so these
3:08
models are degrading right so these
3:08
models are degrading right so these represents your lost revenue because of
3:10
represents your lost revenue because of
3:10
represents your lost revenue because of the model getting degraded. It's a
3:12
the model getting degraded. It's a
3:12
the model getting degraded. It's a manual overhead. You have to actually
3:14
manual overhead. You have to actually
3:14
manual overhead. You have to actually the things which are not able to catch
3:16
the things which are not able to catch
3:16
the things which are not able to catch by the model. You have the manual
3:18
by the model. You have the manual
3:18
by the model. You have the manual overhead there. And of course the poor
3:20
overhead there. And of course the poor
3:20
overhead there. And of course the poor user experience depending on where the
3:22
user experience depending on where the
3:22
user experience depending on where the model is deployed.
3:24
model is deployed.
3:24
model is deployed. So longer we delay the retraining the
3:28
So longer we delay the retraining the
3:28
So longer we delay the retraining the the problem will keep compounding. You
3:29
the problem will keep compounding. You
3:30
the problem will keep compounding. You can imagine the first month what the
3:32
can imagine the first month what the
3:32
can imagine the first month what the problem you're having in couple of
3:33
problem you're having in couple of
3:33
problem you're having in couple of months the problem is just keep
3:35
months the problem is just keep
3:35
months the problem is just keep compounding and problem is becoming
3:36
compounding and problem is becoming
3:36
compounding and problem is becoming intense and intense.
3:42
So talking about the real world
3:44
So talking about the real world
3:44
So talking about the real world consequences like I mean we talked about
3:46
consequences like I mean we talked about
3:46
consequences like I mean we talked about like theoretical the percentage numbers
3:48
like theoretical the percentage numbers
3:48
like theoretical the percentage numbers and all this stuff but what are the real
3:50
and all this stuff but what are the real
3:50
and all this stuff but what are the real world consequences? I'm bringing a few
3:51
world consequences? I'm bringing a few
3:52
world consequences? I'm bringing a few examples in different industries which
3:54
examples in different industries which
3:54
examples in different industries which can help you relate the real world
3:56
can help you relate the real world
3:56
can help you relate the real world consequences of uh the model drift.
3:59
consequences of uh the model drift.
3:59
consequences of uh the model drift. Right? So if you if you just pick up
4:01
Right? So if you if you just pick up
4:01
Right? So if you if you just pick up manu manufacturing industry uh image
4:04
manu manufacturing industry uh image
4:04
manu manufacturing industry uh image classification can lose up to 25 27%
4:08
classification can lose up to 25 27%
4:08
classification can lose up to 25 27% accuracy within 6 months. Like if you
4:11
accuracy within 6 months. Like if you
4:11
accuracy within 6 months. Like if you translate that to that will translate to
4:13
translate that to that will translate to
4:13
translate that to that will translate to roughly $3.2 2 million annual waste and
4:16
roughly $3.2 2 million annual waste and
4:16
roughly $3.2 2 million annual waste and potential like you can have bunch of
4:18
potential like you can have bunch of
4:18
potential like you can have bunch of different product recalls
4:20
different product recalls
4:20
different product recalls in healthcare which is uh critical
4:23
in healthcare which is uh critical
4:23
in healthcare which is uh critical industry even if the model starts with
4:26
industry even if the model starts with
4:26
industry even if the model starts with 94% sensitivity right so that's pretty
4:29
94% sensitivity right so that's pretty
4:29
94% sensitivity right so that's pretty high sensitivity it can drop to 78%.
4:34
high sensitivity it can drop to 78%.
4:34
high sensitivity it can drop to 78%. uh and like effectively missing critical
4:37
uh and like effectively missing critical
4:37
uh and like effectively missing critical diagnosis right so which is not
4:39
diagnosis right so which is not
4:39
diagnosis right so which is not something acceptable
4:41
something acceptable
4:41
something acceptable uh financially like you can think of
4:43
uh financially like you can think of
4:43
uh financially like you can think of like just thinking of the financial
4:45
like just thinking of the financial
4:45
like just thinking of the financial impact of it like every 1% degradation
4:49
impact of it like every 1% degradation
4:49
impact of it like every 1% degradation leads to the cascade right so 1%
4:52
leads to the cascade right so 1%
4:52
leads to the cascade right so 1% degradation can have more like it
4:54
degradation can have more like it
4:54
degradation can have more like it results in more false positives
4:57
results in more false positives
4:57
results in more false positives uh loss sales and you can imagine the
5:00
uh loss sales and you can imagine the
5:00
uh loss sales and you can imagine the rising the operational cost right So
5:03
rising the operational cost right So
5:03
rising the operational cost right So just 1% and imagine like you're losing
5:05
just 1% and imagine like you're losing
5:05
just 1% and imagine like you're losing 25% you're losing uh bunch of 30%
5:08
25% you're losing uh bunch of 30%
5:08
25% you're losing uh bunch of 30% accuracy in these models like 1% can
5:10
accuracy in these models like 1% can
5:10
accuracy in these models like 1% can have such a huge financial impact.
5:15
So now since we have talked about what's
5:19
So now since we have talked about what's
5:19
So now since we have talked about what's a financial implication what's a like we
5:21
a financial implication what's a like we
5:22
a financial implication what's a like we have talked about some of the
5:22
have talked about some of the
5:22
have talked about some of the theoretical numbers on the model uh
5:24
theoretical numbers on the model uh
5:24
theoretical numbers on the model uh drifts that happens like talks about
5:27
drifts that happens like talks about
5:27
drifts that happens like talks about what are the benefits of if you continue
5:30
what are the benefits of if you continue
5:30
what are the benefits of if you continue if you have continuous retraining of
5:31
if you have continuous retraining of
5:31
if you have continuous retraining of your models what are what will be the
5:33
your models what are what will be the
5:33
your models what are what will be the benefit of it I think one is the
5:36
benefit of it I think one is the
5:36
benefit of it I think one is the lifespan you can actually increase the
5:38
lifespan you can actually increase the
5:38
lifespan you can actually increase the model lifespan by 42%.
5:41
model lifespan by 42%.
5:41
model lifespan by 42%. Uh and then the other one is compute
5:44
Uh and then the other one is compute
5:44
Uh and then the other one is compute savings. You can actually have 60% 62%
5:48
savings. You can actually have 60% 62%
5:48
savings. You can actually have 60% 62% saving uh if you basically do regular
5:51
saving uh if you basically do regular
5:52
saving uh if you basically do regular retraining versus if you do the full
5:53
retraining versus if you do the full
5:53
retraining versus if you do the full retraining of the model.
5:56
retraining of the model.
5:56
retraining of the model. The third one is data efficiency like
5:58
The third one is data efficiency like
5:58
The third one is data efficiency like you will have uh we just need like 53%
6:02
you will have uh we just need like 53%
6:02
you will have uh we just need like 53% of the training volume versus like the
6:05
of the training volume versus like the
6:05
of the training volume versus like the full retraining sample if you have to do
6:06
full retraining sample if you have to do
6:06
full retraining sample if you have to do it. The other one which is I think the
6:09
it. The other one which is I think the
6:09
it. The other one which is I think the the benefit of it is like 43% fewer
6:13
the benefit of it is like 43% fewer
6:13
the benefit of it is like 43% fewer false positives
6:15
false positives
6:15
false positives like I mean that is a huge percentage in
6:18
like I mean that is a huge percentage in
6:18
like I mean that is a huge percentage in false positive right it directly
6:21
false positive right it directly
6:21
false positive right it directly improves the trust in the AI system the
6:24
improves the trust in the AI system the
6:24
improves the trust in the AI system the so I mean effectively solution is not
6:26
so I mean effectively solution is not
6:26
so I mean effectively solution is not just technical so it's economical and
6:29
just technical so it's economical and
6:29
just technical so it's economical and strategic for basically the the the
6:31
strategic for basically the the the
6:31
strategic for basically the the the retraining requires have that strategic
6:34
retraining requires have that strategic
6:34
retraining requires have that strategic need that that we we should be focused
6:38
need that that we we should be focused
6:38
need that that we we should be focused Um
6:40
Um
6:40
Um now we have talked about basically the
6:43
now we have talked about basically the
6:43
now we have talked about basically the benefits. Now what is my like what is
6:46
benefits. Now what is my like what is
6:46
benefits. Now what is my like what is the the high level blueprint I'm
6:49
the the high level blueprint I'm
6:49
the the high level blueprint I'm recommending? So high level blueprint is
6:52
recommending? So high level blueprint is
6:52
recommending? So high level blueprint is we I think first is continuously mon
6:55
we I think first is continuously mon
6:55
we I think first is continuously mon monitor model performance like that's
6:57
monitor model performance like that's
6:57
monitor model performance like that's the first thing like if we don't measure
6:59
the first thing like if we don't measure
6:59
the first thing like if we don't measure it we won't know it right so first step
7:01
it we won't know it right so first step
7:01
it we won't know it right so first step is like are you measuring the
7:02
is like are you measuring the
7:02
is like are you measuring the performance of the model so a lot of
7:04
performance of the model so a lot of
7:04
performance of the model so a lot of time happens like the model is deployed
7:06
time happens like the model is deployed
7:06
time happens like the model is deployed we started with 95% accuracy but after
7:09
we started with 95% accuracy but after
7:09
we started with 95% accuracy but after that we have no idea what's the accuracy
7:11
that we have no idea what's the accuracy
7:11
that we have no idea what's the accuracy of the model is
7:13
of the model is
7:13
of the model is then use smart sampling right to collect
7:17
then use smart sampling right to collect
7:17
then use smart sampling right to collect new and informative data for continuous
7:19
new and informative data for continuous
7:19
new and informative data for continuous retrain
7:21
retrain
7:21
retrain and I think the third step is uh use
7:23
and I think the third step is uh use
7:23
and I think the third step is uh use incremental training techniques to
7:25
incremental training techniques to
7:25
incremental training techniques to update the model and very important that
7:27
update the model and very important that
7:27
update the model and very important that we have to res preserve the knowledge
7:29
we have to res preserve the knowledge
7:29
we have to res preserve the knowledge the existing knowledge of that model and
7:32
the existing knowledge of that model and
7:32
the existing knowledge of that model and I I think the final step is through
7:35
I I think the final step is through
7:35
I I think the final step is through validation like before deployment so
7:37
validation like before deployment so
7:37
validation like before deployment so that's very important that once you
7:39
that's very important that once you
7:39
that's very important that once you retrain the model uh make sure that the
7:42
retrain the model uh make sure that the
7:42
retrain the model uh make sure that the model is still behaving the way it
7:44
model is still behaving the way it
7:44
model is still behaving the way it should be behaving
7:45
should be behaving
7:45
should be behaving uh so
7:48
uh so
7:48
uh so so So this keeps the model current
7:50
so So this keeps the model current
7:50
so So this keeps the model current without actually fully retraining the
7:52
without actually fully retraining the
7:52
without actually fully retraining the model. So this is this is the highle
7:55
model. So this is this is the highle
7:55
model. So this is this is the highle four steps you need need to follow.
8:01
Um now I'm talking about like
8:04
Um now I'm talking about like
8:04
Um now I'm talking about like like some of the more details in each of
8:06
like some of the more details in each of
8:06
like some of the more details in each of those steps. One is like how do I know
8:08
those steps. One is like how do I know
8:08
those steps. One is like how do I know when to retrain the model?
8:11
when to retrain the model?
8:11
when to retrain the model? So like we can I I think we can rely on
8:15
So like we can I I think we can rely on
8:16
So like we can I I think we can rely on the fixed calendar like every time like
8:18
the fixed calendar like every time like
8:18
the fixed calendar like every time like every month we try to retrain but the I
8:21
every month we try to retrain but the I
8:21
every month we try to retrain but the I I think the my proposal is focus more on
8:24
I think the my proposal is focus more on
8:24
I think the my proposal is focus more on like that's why we are monitoring the
8:26
like that's why we are monitoring the
8:26
like that's why we are monitoring the model performance so focus more on
8:28
model performance so focus more on
8:28
model performance so focus more on monitoring some of the signals from your
8:30
monitoring some of the signals from your
8:30
monitoring some of the signals from your model like you can actually see if the
8:32
model like you can actually see if the
8:32
model like you can actually see if the model have certain drops in confidence
8:34
model have certain drops in confidence
8:34
model have certain drops in confidence scores uh you're you're noticing the
8:37
scores uh you're you're noticing the
8:37
scores uh you're you're noticing the shifts in prediction distributions
8:41
shifts in prediction distributions
8:41
shifts in prediction distributions um you are noticing drift uh like you
8:43
um you are noticing drift uh like you
8:44
um you are noticing drift uh like you can actually use the couple of
8:45
can actually use the couple of
8:45
can actually use the couple of techniques like KL divergence or ve
8:47
techniques like KL divergence or ve
8:47
techniques like KL divergence or ve sustain distance to see if there's a
8:49
sustain distance to see if there's a
8:50
sustain distance to see if there's a there's a drift detected in your model.
8:52
there's a drift detected in your model.
8:52
there's a drift detected in your model. Uh the other one is canary test
8:54
Uh the other one is canary test
8:54
Uh the other one is canary test failures. You can actually look into
8:55
failures. You can actually look into
8:55
failures. You can actually look into those ones. Right? So these adaptive
8:59
those ones. Right? So these adaptive
8:59
those ones. Right? So these adaptive triggers ensures that retraining happens
9:01
triggers ensures that retraining happens
9:01
triggers ensures that retraining happens before pro performance collapse not
9:04
before pro performance collapse not
9:04
before pro performance collapse not after. Right? And then it's also why
9:06
after. Right? And then it's also why
9:06
after. Right? And then it's also why it's like it gives you a real signal
9:08
it's like it gives you a real signal
9:08
it's like it gives you a real signal when to when to retrain your model
9:10
when to when to retrain your model
9:10
when to when to retrain your model versus you guessing it takes the guess
9:13
versus you guessing it takes the guess
9:13
versus you guessing it takes the guess works out of the when to retrain your
9:15
works out of the when to retrain your
9:15
works out of the when to retrain your model.
9:18
model.
9:18
model. Now going to the next portion
9:21
Now going to the next portion
9:21
Now going to the next portion like the other one is we talked about
9:24
like the other one is we talked about
9:24
like the other one is we talked about like the we need data for retraining
9:26
like the we need data for retraining
9:26
like the we need data for retraining like what are the efficient data
9:27
like what are the efficient data
9:27
like what are the efficient data sampling techniques we can use to
9:29
sampling techniques we can use to
9:30
sampling techniques we can use to retrain our models. So I I think the
9:32
retrain our models. So I I think the
9:32
retrain our models. So I I think the main thing we talked about we want to
9:33
main thing we talked about we want to
9:33
main thing we talked about we want to minimize the amount of data we need
9:35
minimize the amount of data we need
9:35
minimize the amount of data we need right so we can actually focus on these
9:37
right so we can actually focus on these
9:37
right so we can actually focus on these four techniques. One is uh uncertaintity
9:40
four techniques. One is uh uncertaintity
9:40
four techniques. One is uh uncertaintity sampling. So we targets the input that
9:42
sampling. So we targets the input that
9:42
sampling. So we targets the input that model struggles with right. The other
9:45
model struggles with right. The other
9:45
model struggles with right. The other one is diversity sampling. So instead of
9:48
one is diversity sampling. So instead of
9:48
one is diversity sampling. So instead of having lot of repetitions we avoid
9:49
having lot of repetitions we avoid
9:50
having lot of repetitions we avoid repetitions and we try to cover the
9:51
repetitions and we try to cover the
9:51
repetitions and we try to cover the entire feature space. Uh the other one
9:54
entire feature space. Uh the other one
9:54
entire feature space. Uh the other one is adversarial sampling. So add the edge
9:57
is adversarial sampling. So add the edge
9:57
is adversarial sampling. So add the edge cases to strengthen the robot robustness
10:00
cases to strengthen the robot robustness
10:00
cases to strengthen the robot robustness of the models and I think last but not
10:03
of the models and I think last but not
10:03
of the models and I think last but not least is active learning. Uh basically
10:06
least is active learning. Uh basically
10:06
least is active learning. Uh basically selectively involves human annotators.
10:08
selectively involves human annotators.
10:08
selectively involves human annotators. Um I mean so all these techniques can
10:11
Um I mean so all these techniques can
10:11
Um I mean so all these techniques can help us avoid reduce the labeling cost
10:14
help us avoid reduce the labeling cost
10:14
help us avoid reduce the labeling cost and uh retraining cycle significantly
10:17
and uh retraining cycle significantly
10:17
and uh retraining cycle significantly right. So I mean we that's a that's a
10:19
right. So I mean we that's a that's a
10:19
right. So I mean we that's a that's a goal of the retraining and that's one of
10:20
goal of the retraining and that's one of
10:20
goal of the retraining and that's one of the advantage of the retraining on a
10:22
the advantage of the retraining on a
10:22
the advantage of the retraining on a regular basis.
10:28
So
10:30
So
10:30
So one major risk so we talked about like
10:33
one major risk so we talked about like
10:33
one major risk so we talked about like different uh facets of uh retraining how
10:36
different uh facets of uh retraining how
10:36
different uh facets of uh retraining how to focus on those one triggers and
10:39
to focus on those one triggers and
10:39
to focus on those one triggers and basically efficient sampling.
10:41
basically efficient sampling.
10:41
basically efficient sampling. One major risk with continuous learning
10:43
One major risk with continuous learning
10:43
One major risk with continuous learning is catastrophic forgetting. So so where
10:47
is catastrophic forgetting. So so where
10:47
is catastrophic forgetting. So so where what I mean by that is where the model
10:49
what I mean by that is where the model
10:49
what I mean by that is where the model forgets what it used to know. to prevent
10:52
forgets what it used to know. to prevent
10:52
forgets what it used to know. to prevent this. I think uh like I'm suggesting a
10:54
this. I think uh like I'm suggesting a
10:54
this. I think uh like I'm suggesting a couple of things to to make sure that we
10:57
couple of things to to make sure that we
10:57
couple of things to to make sure that we prevent this one. This one is very
10:58
prevent this one. This one is very
10:58
prevent this one. This one is very important that models learn new things
11:01
important that models learn new things
11:01
important that models learn new things versus forgetting the the which which is
11:03
versus forgetting the the which which is
11:03
versus forgetting the the which which is already new. So one is elast elastic
11:05
already new. So one is elast elastic
11:05
already new. So one is elast elastic weight consolidations. Uh this will help
11:08
weight consolidations. Uh this will help
11:08
weight consolidations. Uh this will help to preserve important weights. Uh the
11:10
to preserve important weights. Uh the
11:10
to preserve important weights. Uh the other one is knowledge distillations. Uh
11:13
other one is knowledge distillations. Uh
11:13
other one is knowledge distillations. Uh so we want to make sure that we transfer
11:15
so we want to make sure that we transfer
11:15
so we want to make sure that we transfer the wisdom from old to new models.
11:19
the wisdom from old to new models.
11:19
the wisdom from old to new models. And the third one is replay buffers for
11:22
And the third one is replay buffers for
11:22
And the third one is replay buffers for historical data. Right? So we actually
11:24
historical data. Right? So we actually
11:24
historical data. Right? So we actually use replay buffers for historical data
11:26
use replay buffers for historical data
11:26
use replay buffers for historical data and gradient regularization
11:30
and gradient regularization
11:30
and gradient regularization to avoid harmful updates. So the the
11:33
to avoid harmful updates. So the the
11:33
to avoid harmful updates. So the the whole goal of this exercise is uh
11:36
whole goal of this exercise is uh
11:36
whole goal of this exercise is uh helping the model learn something new
11:39
helping the model learn something new
11:39
helping the model learn something new without forgetting which is already
11:41
without forgetting which is already
11:41
without forgetting which is already mastered which is very critical. Right?
11:43
mastered which is very critical. Right?
11:43
mastered which is very critical. Right? So we want to basically just focus on
11:46
So we want to basically just focus on
11:46
So we want to basically just focus on the retraining of new things and making
11:48
the retraining of new things and making
11:48
the retraining of new things and making sure that previous knowledge is
11:50
sure that previous knowledge is
11:50
sure that previous knowledge is preserved.
11:56
Okay. So couple of key takeaways.
12:00
Okay. So couple of key takeaways.
12:00
Okay. So couple of key takeaways. Um I think to sum it up here are four
12:05
Um I think to sum it up here are four
12:05
Um I think to sum it up here are four four key takeaways right. First is like
12:10
four key takeaways right. First is like
12:10
four key takeaways right. First is like use drift detection methods
12:13
use drift detection methods
12:13
use drift detection methods using KL divergence and variance metrics
12:15
using KL divergence and variance metrics
12:16
using KL divergence and variance metrics and other methods to see if there is
12:18
and other methods to see if there is
12:18
and other methods to see if there is model is drifting right so it goes back
12:21
model is drifting right so it goes back
12:21
model is drifting right so it goes back to measuring right so measuring is a
12:23
to measuring right so measuring is a
12:23
to measuring right so measuring is a major focus uh be proactive in figuring
12:26
major focus uh be proactive in figuring
12:26
major focus uh be proactive in figuring out if any drift is happening for your
12:28
out if any drift is happening for your
12:28
out if any drift is happening for your model or not
12:31
model or not
12:31
model or not um and then within your uh ML life cycle
12:36
um and then within your uh ML life cycle
12:36
um and then within your uh ML life cycle build adaptive retraining workflows
12:39
build adaptive retraining workflows
12:39
build adaptive retraining workflows with sampling and incremental updates,
12:42
with sampling and incremental updates,
12:42
with sampling and incremental updates, right? So you want to make sure that
12:43
right? So you want to make sure that
12:43
right? So you want to make sure that you're sampling your you build the
12:45
you're sampling your you build the
12:45
you're sampling your you build the efficient sampling techniques so that
12:47
efficient sampling techniques so that
12:47
efficient sampling techniques so that the sampling volume is low and you do
12:51
the sampling volume is low and you do
12:51
the sampling volume is low and you do incremental updates. So you build like
12:53
incremental updates. So you build like
12:53
incremental updates. So you build like the you're able to quickly update your
12:56
the you're able to quickly update your
12:56
the you're able to quickly update your model.
12:57
model.
12:57
model. And then the third one is you validate
13:01
And then the third one is you validate
13:01
And then the third one is you validate to make sure that the new model is
13:04
to make sure that the new model is
13:04
to make sure that the new model is learning new things with the old
13:07
learning new things with the old
13:07
learning new things with the old knowledge preserved, right? So you want
13:09
knowledge preserved, right? So you want
13:09
knowledge preserved, right? So you want to make sure you want to be super
13:11
to make sure you want to be super
13:11
to make sure you want to be super critical that you're not losing the old
13:13
critical that you're not losing the old
13:13
critical that you're not losing the old knowledge of your models and you're
13:14
knowledge of your models and you're
13:14
knowledge of your models and you're learning the new things in the model. So
13:16
learning the new things in the model. So
13:16
learning the new things in the model. So uh validation is super important and it
13:18
uh validation is super important and it
13:18
uh validation is super important and it has these dual purposes, right? So one
13:20
has these dual purposes, right? So one
13:20
has these dual purposes, right? So one is your model can you can validate your
13:23
is your model can you can validate your
13:23
is your model can you can validate your model, learn new things. The other one
13:24
model, learn new things. The other one
13:24
model, learn new things. The other one is you didn't lose your old knowledge of
13:26
is you didn't lose your old knowledge of
13:26
is you didn't lose your old knowledge of the model.
13:28
the model.
13:28
the model. Uh I think the the fourth thing which is
13:30
Uh I think the the fourth thing which is
13:30
Uh I think the the fourth thing which is very important is which helps you give
13:33
very important is which helps you give
13:33
very important is which helps you give the sense of how critical retraining
13:36
the sense of how critical retraining
13:36
the sense of how critical retraining regular retraining of the model is that
13:38
regular retraining of the model is that
13:38
regular retraining of the model is that you track the total cost of the
13:40
you track the total cost of the
13:40
you track the total cost of the ownership.
13:42
ownership.
13:42
ownership. uh you have to like once you employ
13:44
uh you have to like once you employ
13:44
uh you have to like once you employ retraining you will see that you're uh
13:47
retraining you will see that you're uh
13:47
retraining you will see that you're uh like how much your return you want to
13:49
like how much your return you want to
13:49
like how much your return you want to calculate your return in how much fewer
13:51
calculate your return in how much fewer
13:51
calculate your return in how much fewer errors you are seeing right or false
13:53
errors you are seeing right or false
13:53
errors you are seeing right or false positive you're seeing with the regular
13:54
positive you're seeing with the regular
13:54
positive you're seeing with the regular retraining
13:56
retraining
13:56
retraining uh how what is the lifespan of the model
13:59
uh how what is the lifespan of the model
13:59
uh how what is the lifespan of the model is it increasing or not uh how much
14:02
is it increasing or not uh how much
14:02
is it increasing or not uh how much compute is actually how efficient the
14:04
compute is actually how efficient the
14:04
compute is actually how efficient the compute is becoming is it going down uh
14:07
compute is becoming is it going down uh
14:07
compute is becoming is it going down uh is it increasing right so like I mean
14:09
is it increasing right so like I mean
14:09
is it increasing right so like I mean you will see the the cost coming down
14:11
you will see the the cost coming down
14:11
you will see the the cost coming down your compute was coming down as you
14:13
your compute was coming down as you
14:13
your compute was coming down as you retrain on a regular basis. So I think
14:16
retrain on a regular basis. So I think
14:16
retrain on a regular basis. So I think tracking that is very important and
14:18
tracking that is very important and
14:18
tracking that is very important and especially with the existing system you
14:19
especially with the existing system you
14:19
especially with the existing system you have or existing life cycle ML life
14:21
have or existing life cycle ML life
14:21
have or existing life cycle ML life cycles you have right. So start tracking
14:23
cycles you have right. So start tracking
14:23
cycles you have right. So start tracking it right away before you employ the
14:26
it right away before you employ the
14:26
it right away before you employ the retraining strategies.
14:32
Um the last thank you for time today. Uh
14:37
Um the last thank you for time today. Uh
14:37
Um the last thank you for time today. Uh I hope this session offered you
14:39
I hope this session offered you
14:39
I hope this session offered you practical insights into building
14:41
practical insights into building
14:41
practical insights into building resilient and sustainable AI system for
14:43
resilient and sustainable AI system for
14:43
resilient and sustainable AI system for the real use cases. I mean this project
14:45
the real use cases. I mean this project
14:46
the real use cases. I mean this project was really close to me and I'm excited
14:47
was really close to me and I'm excited
14:48
was really close to me and I'm excited to see more organizations
14:50
to see more organizations
14:50
to see more organizations and uh individuals and institutes employ
14:55
and uh individuals and institutes employ
14:55
and uh individuals and institutes employ embrace this continuous learning as part
14:57
embrace this continuous learning as part
14:57
embrace this continuous learning as part of their uh production AI strategy.
14:59
of their uh production AI strategy.
15:00
of their uh production AI strategy. Thank you.
#Machine Learning & Artificial Intelligence