Auto retrain ML models is not the question

MLOps.community · Intermediate ·📐 ML Fundamentals ·6y ago

Key Takeaways

The video discusses the importance of auto retraining machine learning models, with Shubhi Jain from SurveyMonkey sharing their approach to enabling auto retraining for certain use cases, while also having a backup process in place to retrain models if needed.

Full Transcript

I just wanted to ask one more about that Auto retrain because last week when Phil was on he made the case for this idea of good practice being just to Auto retrain but he wasn't sure it was so worth it in some - most use cases or what he had seen with certain companies it wasn't needed do you have Auto retraining going on or is that something that is not necessary how does that work with you guys so I agree with Phil as well I think there are some use cases where you do not need it and I think that's the same thing for us there are some use cases where we do enable it in some use cases where we do not however we whether we enable auto every trainer or not is one thing but still being able to auto retrain is I think another thing all of our models if we need to retrain our models and fresh data we can we can just click a button and it'll go retrain and the reason for that is because if anything were to get corrupted if anything were we realize ago we lost our model artifact or whole like repository of our marble artifacts went down we know we can go and do it again and we're confident in that and that is just know our backup processes we want to make sure everything is good there and I think often we think like oh everything's in the cloud it's safe but you know s3 goes down to AWS goes down to and we like to just have our backups there for those keys for those you know that less than 1% of the time when things go wrong and it also allows us to have a level of accountability of making sure like oh we did what we did during training and we can reproduce it and then reproducibility aspect make sure that if someone is leaving the company and someone you is joining it makes sense to them what went on here and they can get the exact same model artifact once again so it can be used in a case so a couple of different smaller reasons there that forced us to have to retrain available still there but it's still only enabled for certain use cases where we see data drift occurring or where we see that we are our performance is varying over time yeah

Original Description

MLOps Community Meetup #4 In the 4th online meetup for our MLOps.community We spoke with Shubhi Jain, Machine Learning Engineer and all-around great guy! In this Clip he talks about why SurveyMonkey decided to have the ability to auto retrain Machine Learning Models. This is an excerpt taken from the longer conversation that can be found here: https://youtu.be/oq1g4s2dUHE Every organization is leveraging machine learning (ML) to provide increasing value to their customers and understand their business. You may have created models too. But, how do you scale this process now? In this case study, we looked at how to pinpoint inefficiencies in your ML data flow, how SurveyMonkey tackled this, and how to make your data more usable to accelerate ML model development. Shubhi Jain is a machine learning engineer at SurveyMonkey where he develops and implements machine learning systems for its products and teams. Occasionally, he’ll create YouTube videos about Machine Learning in collaboration with Springboard, an e-learning platform. He’s always excited to bring his expertise and passion for Data and AI systems to the rest of the industry. In his free time, Shubhi likes hiking with his dog and accelerating his hearing loss at live music shows. This was a virtual fireside chat between Shubhi Jain, Demetrios Brinkmann and the MLOps community. Relevant links can be found below. Join our MLOps slack community: https://bit.ly/3aOTwgR Connect with Demetrios on LinkedIn: https://www.linkedin.com/in/dpbrinkm/ Connect with Shubhi Jain on Linkedin: https://www.linkedin.com/in/shubhankarjain/
Watch on YouTube ↗ (saves to browser)
Sign in to unlock AI tutor explanation · ⚡30

Playlist

Uploads from MLOps.community · MLOps.community · 28 of 60

← Previous Next →
1 Our 1st MLOps Meetup // Luke Marsden // MLOps Meetup #1
Our 1st MLOps Meetup // Luke Marsden // MLOps Meetup #1
MLOps.community
 2 Remote Collaboration as a Data Scientist
Remote Collaboration as a Data Scientist
MLOps.community
 3 MLOps Manifesto with Luke Marsden from Dotscience
MLOps Manifesto with Luke Marsden from Dotscience
MLOps.community
 4 MLOps lifecycle description
MLOps lifecycle description
MLOps.community
 5 What Does Best in Class AI/ML Governance Look Like in Fin Services? // Charles Radclyffe // MLOps #2
What Does Best in Class AI/ML Governance Look Like in Fin Services? // Charles Radclyffe // MLOps #2
MLOps.community
 6 Life purpose and too many spreadsheets
Life purpose and too many spreadsheets
MLOps.community
 7 Explainability, Black boxes and EU white paper on reproducibility
Explainability, Black boxes and EU white paper on reproducibility
MLOps.community
 8 Hierarchy of Machine Learning Needs // Phil Winder // MLOps Meetup #3
Hierarchy of Machine Learning Needs // Phil Winder // MLOps Meetup #3
MLOps.community
 9 Automatically Retrain Machine Learning Models? Are best practices worth it?
Automatically Retrain Machine Learning Models? Are best practices worth it?
MLOps.community
 10 Building an MLOps Team? Key ideas to keep in mind
Building an MLOps Team? Key ideas to keep in mind
MLOps.community
 11 Hierarchy of MLOps Needs
Hierarchy of MLOps Needs
MLOps.community
 12 Bare necessities for getting an ML model into production
Bare necessities for getting an ML model into production
MLOps.community
 13 MLOps and Monitoring
MLOps and Monitoring
MLOps.community
 14 How Phil Winder got into Data Science and Software Engineering
How Phil Winder got into Data Science and Software Engineering
MLOps.community
 15 Provenance and Reproducibility in Machine Learning; what is it and why you need it?
Provenance and Reproducibility in Machine Learning; what is it and why you need it?
MLOps.community
 16 Friction Between Data Scientists and Software Engineers
Friction Between Data Scientists and Software Engineers
MLOps.community
 17 MLOps Problems in different size companies
MLOps Problems in different size companies
MLOps.community
 18 ML tooling in large companies
ML tooling in large companies
MLOps.community
 19 ML Platforms - The build vs buy question
ML Platforms - The build vs buy question
MLOps.community
 20 ML Services Gateway at SurveyMonkey
ML Services Gateway at SurveyMonkey
MLOps.community
 21 Message buses, Async and sync architecture
Message buses, Async and sync architecture
MLOps.community
 22 MLOps #4: Shubhi Jain - Building an ML Platform @SurveyMonkey
MLOps #4: Shubhi Jain - Building an ML Platform @SurveyMonkey
MLOps.community
 23 Hybrid Data Science Teams @SurveyMonkey
Hybrid Data Science Teams @SurveyMonkey
MLOps.community
 24 How do you handle ML version control at SurveyMonkey
How do you handle ML version control at SurveyMonkey
MLOps.community
 25 Doing ML with Personal Information
Doing ML with Personal Information
MLOps.community
 26 Evolution of the ML feature store @SurveyMonkey
Evolution of the ML feature store @SurveyMonkey
MLOps.community
 27 Developing a Machine Learning Feature Store
Developing a Machine Learning Feature Store
MLOps.community
 ▶ Auto retrain ML models is not the question
Auto retrain ML models is not the question
MLOps.community
 29 3 key parts to Machine Learning monitoring
3 key parts to Machine Learning monitoring
MLOps.community
 30 MLOps Meetup #6: Mid-Scale Production Feature Engineering with Dr. Venkata Pingali
MLOps Meetup #6: Mid-Scale Production Feature Engineering with Dr. Venkata Pingali
MLOps.community
 31 MLOps meetup #5 High Stakes ML: Active Failures, Latent Factors with Flavio Clesio
MLOps meetup #5 High Stakes ML: Active Failures, Latent Factors with Flavio Clesio
MLOps.community
 32 MLOps: Airflow Pros and Cons
MLOps: Airflow Pros and Cons
MLOps.community
 33 Specific challenges in Machine Learning
Specific challenges in Machine Learning
MLOps.community
 34 Current State Of Machine Learning
Current State Of Machine Learning
MLOps.community
 35 Humans in the Loop are a defining factor in Machine Learning
Humans in the Loop are a defining factor in Machine Learning
MLOps.community
 36 Learning from real life Machine Learning failures
Learning from real life Machine Learning failures
MLOps.community
 37 Survivorship Bias in machine learning tutorials
Survivorship Bias in machine learning tutorials
MLOps.community
 38 Swiss Cheese model in Machine Learning
Swiss Cheese model in Machine Learning
MLOps.community
 39 Resume driven development in Machine learning & software engineering
Resume driven development in Machine learning & software engineering
MLOps.community
 40 Who has the highest standards in ML?
Who has the highest standards in ML?
MLOps.community
 41 Venkata Pingali of Scribble Data Thoughts on the Current State of Machine Learning
Venkata Pingali of Scribble Data Thoughts on the Current State of Machine Learning
MLOps.community
 42 Dependable data and being able to Trust in your Data with Venkata Pengali of Scribble Data
Dependable data and being able to Trust in your Data with Venkata Pengali of Scribble Data
MLOps.community
 43 Speed, Trust, Evolution and Scale in MLOps
Speed, Trust, Evolution and Scale in MLOps
MLOps.community
 44 More difficult transition for data scientists to become ML engineers
More difficult transition for data scientists to become ML engineers
MLOps.community
 45 How many models in prod til I need a dedicated ML platform?
How many models in prod til I need a dedicated ML platform?
MLOps.community
 46 Deeper thinking from data scientists around platform blackholes
Deeper thinking from data scientists around platform blackholes
MLOps.community
 47 Checkpointing, metadata, and confidence in your data
Checkpointing, metadata, and confidence in your data
MLOps.community
 48 Adjacent usecases and multistep feature engineering
Adjacent usecases and multistep feature engineering
MLOps.community
 49 Standardization of Machine Learning tools like in Software Engineering with Venkata Pingali
Standardization of Machine Learning tools like in Software Engineering with Venkata Pingali
MLOps.community
 50 Reproducability flaws in end to end Machine Learning debugging
Reproducability flaws in end to end Machine Learning debugging
MLOps.community
 51 3rd wave of data scientists
3rd wave of data scientists
MLOps.community
 52 MLOps meetup #7 Alex Spanos // TrueLayer 's MLOps Pipeline
MLOps meetup #7 Alex Spanos // TrueLayer 's MLOps Pipeline
MLOps.community
 53 MLOps Meetup #8 Optimizing Your ML Workflow with Kubeflow 1.0
MLOps Meetup #8 Optimizing Your ML Workflow with Kubeflow 1.0
MLOps.community
 54 Are Kubeflow and Airflow complementary?
Are Kubeflow and Airflow complementary?
MLOps.community
 55 Why Kubeflow gained so much traction=open community
Why Kubeflow gained so much traction=open community
MLOps.community
 56 Who decides the dirrection of Kubeflow
Who decides the dirrection of Kubeflow
MLOps.community
 57 What do Kubeflow and Arrikto do and how do they work together?
What do Kubeflow and Arrikto do and how do they work together?
MLOps.community
 58 Versioning your ML steps with Kubeflow
Versioning your ML steps with Kubeflow
MLOps.community
 59 Machine Learning Lifecycles//Perception vs Reality
Machine Learning Lifecycles//Perception vs Reality
MLOps.community
 60 Kubeflow vs SageMaker in Machine Learning
Kubeflow vs SageMaker in Machine Learning
MLOps.community

The video highlights the importance of auto retraining machine learning models, particularly in cases where data drift occurs or model performance varies over time. SurveyMonkey's approach to enabling auto retraining for certain use cases is discussed, along with their backup process for model artifacts.

Key Takeaways
  1. Determine if auto retraining is necessary for a particular use case
  2. Implement auto retraining for ML models
  3. Develop backup processes for model artifacts
  4. Monitor model performance over time
  5. Enable auto retraining for use cases with data drift or varying performance
💡 Having a backup process in place to retrain models if needed is crucial, even if auto retraining is not enabled for all use cases.

Related AI Lessons

⚡
Beyond the Elephant: On Manifolds, Projections, and the Hidden Assumptions of Neural Geometry
Learn how neural geometry relies on manifolds, projections, and hidden assumptions to understand complex data, and why it matters for AI development
Medium · AI
 ⚡
Beyond the Elephant: On Manifolds, Projections, and the Hidden Assumptions of Neural Geometry
Learn how neural geometry relies on manifolds, projections, and hidden assumptions to understand complex data, and why it matters for advancing AI research
Medium · Data Science
 ⚡
Beyond the Elephant: On Manifolds, Projections, and the Hidden Assumptions of Neural Geometry
Explore the geometric assumptions underlying neural networks and their implications on manifold learning and projections
Medium · Deep Learning
 ⚡
Beyond the Elephant: On Manifolds, Projections, and the Hidden Assumptions of Neural Geometry
Learn about the hidden assumptions of neural geometry and how manifolds and projections impact neural network performance
Medium · LLM
Up next
Machine Learning Project for Final Year Students | ML Project Idea @FameWorldEducationalHub
FAME WORLD EDUCATIONAL HUB
Watch →