Short AI Primer

Transcript

1. AI Primer May 2018 Brief Note
2. Core of artificial intelligence 2 Prediction Using information we have, get information we do not have Predict demand for yogurt in a store Predict onset of diabetic retinopathy from fondus image Predict machine failure in advance using anomaly detection Predict apparel type, classify the image accordingly Predict fraud in a credit card transaction
3. What’s driving AI adoption? 3 Declining cost of prediction As cost of making a prediction drops, it will find more takers and more use-cases emerge Drop in compute and memory prices Widespread adoption of computers Drop in price prediction Conventional problem recast as prediction problem, wider adoption e.g. retail self-checkout GPUs, new techniques that need less training data, open source frameworks, rising talent make AI more accessible
4. 4 Primary technique at play Neural Network Supervised Learning Supervised learning uses a large amount of labeled data to learn features and then builds a core classifier or predictor Labeled Data Neural Network Prediction Validation Data Simplified AI Workflow
5. 5 Companies in AI Landscape Vertical AI Horizontal AI Vertical AI companies are full-stack product companies that solve one specific problem whereas horizontal AI companies range from AI infrastructure to commodity APIs Non-invasive breast cancer detection Deep Learning based demand planning for fresh foods Computer vision for insurance claims management Online payment fraud detection Vertical AI Companies Machine learning APIs for developers AI infrastructure for deployment Kubernetes for training and re-producible experiments Cloud infrastructure for AI Horizontal AI Companies
6. 6 Vertical AI Companies Full-stack point solutions, powered by AI Source: Shivon Zilis Bloomberg Beta
7. 7 Horizontal AI Companies - It’s a spectrum! MLaaS Companies - APIs Custom MLaaS APIs Bring Data-Take Model Managed AI Infrastructure Cloud Providers AI Tools & Infrastructure Model Positioning Don’t need data scientists - developers can use it directly with no additional overhead Need a full data science team, data scientists, data engineers and devops engineers Only data scientists are enough to build and deploy models in production Target Users Developers Team: Data scientist + Data Engineers + Devops Engineers Data Scientists
8. 8 Moat is in proprietary data Labeled data is the new new oil Access to and ownership of proprietary data is main source of competitive advantage. Enterprises will prefer to retain data and build ML capabilities in-house Labeled Data Talent Infrastructure e.g. GPUs, Docker
9. 9 Why build machine capabilities in-house? MLaaS Companies - APIs Custom MLaaS APIs Bring Data-Take Model Managed AI Infrastructure Cloud Providers AI Tools & Infrastructure Data is shared with partners Data remains in-house Easier to get started, with low resource and capital requirement More difficult, but a strategy set for pay-off handsomely in the long run Implementation difficulty along the spectrum Low High
10. 10 AI adoption trends - current and future MLaaS Companies - APIs Custom MLaaS APIs Bring Data-Take Model Managed AI Infrastructure Cloud Providers AI Tools & Infrastructure Current adoption curve Future adoption curve
11. Further Reading 11 1. Prediction, Judgment and Complexity: A Theory of Decision Making and Artificial Intelligence 2. Vertical AI Startups: Solving Industry-specific Problems by Combining AI and Subject Matter Expertise 3. Building an AI Startup: Realities & Tactics 4. The Current State of Machine Intelligence 3.0
12. Contact 12 rorodata Raghavendra Badaskar - raghav@rorodata.com