2. ct@askatlas.ai2
ABOUT
• Founder and CEO at Atlas AI
• Launch Monday!
• Over 20 Years in Technology
• Past 2 Years focused on AI Technologies
• Author:
– The AI Revolution, The Future of
Profit
– Marketing AI: 90 Days to AI-Driven
Marketing (Next Month!)
7. ct@askatlas.ai7
FEATUREENGINEERINGEXAMPLES
• Scaling: Transforms a number to remove anomalies in the data
• Binning: Grouping similar data ranges together
• Encoding: Many models require categorical data to be transformed into numbers
• Cleansing: Removing Null values and replacing with zeros, averages, or medians
• Choosing which features are predictive
10. ct@askatlas.ai10
DESIREDSOLUTION
• Model Hosted in the Cloud
• Persistence/Availability
• APIAvailable for Consumption
• Traceable
• Independent
API
Storage
Deployment
LogsPredictive
Result
31. ct@askatlas.ai31
SAGEMAKERSUMMARY
• Notebooks have support for all the major ML & DL frameworks
• Versions of data/models are stored in S3
• You’ll need to create your own checkpointing
• Can be hard to debug
• Don’t forget to turn off instances!
33. ct@askatlas.ai33
TAKEAWAYQUESTIONS
• How does your ML pipeline handle checkpointing?
– Google Big Query you have to check a checkbox for long-running transactions
• How do you engineer features for your organization
– On your local machine?
– Does your API engineer the features from source data?
– Creating a database of shared features and ‘continuous feature engineering’
34. ct@askatlas.ai34
TAKEAWAYQUESTIONS(CONT.)
• How do you store history on the four components of a model version?
– Data, Code, Features, Hyper Parameters, and Training Dataset
• How do you monitor model decay?
– Does your solution monitor for false positive/negative results?
Ad-hoc: you create and deploy your own servers
Dev-ops: you hand off your code to a team of engineers to deploy
Jointly manage: preset deployment configuration, configuration changes need review
Deployment platforms: datarobot, alteryx, ayasdi