Building Efficient ML Pipelines and Responsible AI Solutions Adi Polak Microsoft @adipolak
Trust
• LET’S START FROM THE BEGINNING. What happens when we get raw data? @adipolak
@adipolak
ML Process / Life Cycle 1 Gather Data 2 Feature Extract, Clean and Normalize 3 Select algorithm Repeat! 4 Evaluate model 5 Data/Insights visualization @adipolak
@adipolak
But in real life: Accuracy < 0.5 ROC curve ☹ @adipolak
Aim for high Accuracy @adipolak
What can you do? Automate! @adipolak
HOW? Pipelines!
What are pipelines?
Big Data/ ML Pipelines Visualize Azure Machine Learning @adipolak
Demo Apache Spark ML Pipelines @adipolak
@adipolak Stepan Pushkarev, CTO, Hydrosphere.io
Big Data/ ML Pipelines Visualize Azure Machine Learning Azure Machine Learning @adipolak
High accuracy! But, at what cost? @adipolak
false positives @adipolak
Human centric Responsible AI @adipolak
Pixabay
Our updated goals: Lawful Ethical Robust @adipolak
ML is a black box ML algorithm Training: Model Data Data Model Testing/Prediction: Prediction @adipolak
Big Data/ ML Pipelines Visualize Azure Machine Learning Azure Machine Learning @adipolak
Big Data/ ML Pipelines Visualize Azure Check and transform the Machine Learning data s r e Balance the data n i a l p x Visualize the model E @adipolak
How ONLINE FREE INVESTED 1B$ IN Microsoft HIGH QUALITY OPEN AI support COURSES Responsible AI 115M$ GRANT FOR OPEN SOURCE AI FOR GOOD @adipolak aka.ms/free-responsible-ai-cour aka.ms/ml-interpretability-to se ol
@adipolak aka.ms/ml-interpretability-to ol
aka.ms/ml-interpretability-to ol
Azure Cognitive Services aka.ms/AA6kex s @adipolak
Tools Spark Spark ML Streaming MLflow Spark SQL @adipolak
But in real life: Accuracy < 0.5 ROC curve ☹
Demo Apache Spark ML Pipelines with Cognitive Services @adipolak
Demo @adipolak
You are only Good as your Data is Use explainers Understand your data @adipolak
Learn more ! Thank you ! aka.ms/free-responsible-ai-course aka.ms/twitter_sentiment_analysis aka.ms/ml-interpretability-tool aka.ms/ai-for-good-grant @adipolak
What is Machine learning • Lifecycle: • Gather data • Data preparation – clean it • Data wrangling • Data analysis • Feature extraction • Train model • Test model • Deployment
Recommend
More recommend