Principles of Good Machine Learning Systems Design

Key Notes

• ML in research vs Production

• Batched vs Single call

• ML Production myths
• Deployment is hard

• Build structure around maintenance/monitor

• Software decays over time

• ROI takes time
• Iterative process

• ML Adoption

Keep Thinking!!!

Weekend Lessons - Humans & AI in Financial Services - The Future - Samik Chandarana | ODSC Europe 2019

Key Notes

• Being a thought leaders matters
• Connect and finite transformative ideas

• Why we do, when we do
• Intelligently automating
• Age-old process

• Business Aligned outcomes
• Long term thinking - research projects
• Dream lot / Dream big

• AI for trading
• Suggest probable matches
• Portfolio management
• Reconciliation using AI

• Data Remediation - People to cleanup data
• Libor Transition - Financial crimes
• Digitize information
• Market Surveillance Toolkit
• Get data in order
• Organize information in good dashboard
• Email - Intent extract - Get info - Manual Validation - Notify
• AI is young science
• NLP, Vision, Search

• Finding money graph/ traffic
• Routes used

Keep Thinking!!!

Weekend Lessons - ML in Production: Serverless and Painless - Oliver Gindele, PhD | ODSC Europe 2019

Slides - Link 

Key Notes

• Discovery Phase
• Model building Phase
• Deploy Phase
• Discover - Develop - Deploy

• Develop - Iterate - Validate - Deploy
• Monitoring / Reliability / Highly Available Services

• Scan and Fetch details
• Fetch information / Details from product
• Prototype built on core ios

• Platform independent architecture
• Robust pipeline
• Build new models
• Mobilenet Transfer Learning architecture

• Object texture / size / shape
• Handmade products with variants
• Product shape degrades over time
• Lighting/occlusion challenges
• Data Augmentation / Randomization
• Pretrained weights on imagenet
• Model F1 Score Accuracy
• Model Pipeline implementation

• Building cross functional teams
• Devops Tools mastering
• ML Tools

• Custom and Tailored solutions from GCP
• Own custom solution to productization

• Model Factory
• Preprocessing
• Automation / Components
• Model Validation
• Model Versioning

• Tensorflowlite for cross device deployment

• Small model footprint
• Storage
• Power consumption of APP
• Model in Cloud

• Image in object Storage

• Data Augmentation Pipeline
• Playing with color distribution / Saturation

• Python pipeline
• Apache beam
• Streaming API
• Batch processing
• Cloud dataflow execution engine
• Automated Training Pipeline
• Automated Deployment Pipeline

• Metrics storage in bigquery
• Save Every image in call

• Lower precision on Inference (Int8)
• During training, we can keep precision

• Google Cloud Composer
• Run Airflow for VM
• Build custom workflows
• Airflow to orchestrate data tasks
• Write each tasks in python
• Monitoring / Alerting
• Failure handing / Notification

Happy Learning!!!