Responsible AI for Engineers

Explore responsible AI practices for engineers in this GOTO Copenhagen 2019 conference talk. Delve into the challenges and opportunities presented by AI advancements, including bias, adversarial attacks, and unintended societal impacts. Learn about Google's approach to responsible AI through principles, governance, and design practices. Examine techniques, research findings, and open challenges in areas such as unintended consequences, model understanding, secondary metrics, and engineering objectives. Gain insights from a case study demonstrating how these considerations come together in practice. Discover emerging approaches for aligning machine learning with human values and understand why responsibility is crucial in meeting stakeholder expectations. Benefit from the speaker's expertise as Technical Director for Applied Artificial Intelligence at Google Cloud, covering topics like ImageNet, Google Search, Diabetic Retinopathy, and Google Assistant. Explore concepts such as disproportionate performance, filter bubbles, data cards, model interpretation, and visualization tools. Gain practical knowledge on addressing fairness in AI systems, including strategies for data collection, loss function adjustments, and overall system effectiveness evaluation.

Intro
Rons background
Overview
Imagenet
Google Search
Diabetic Retinopathy
Google Assistant
Google Cloud
Problems with AI
disproportionate performance
Compass
Microsoft Twitter
Filter Bubbles
Amazon Search Algorithm
Principles
Digital Wellbeing
Machine Learning AI
Data Cards
Model Cards
Model Interpretation
Local Interpretation
Shapley Values
Integrated gradients
Gradients
Examples
Visualization Tools
Multiple metrics
Secondary models
Over optimizing
Case study fairness
How does prediction change
How effective is this system
What if we generalize
Collecting more data
Changing the loss function
The net effect
Lessons
AI Guidebook
AI Ethics Guidelines