YoVDO

All of Our ML Ideas Are Bad - and We Should Feel Bad

Offered By: USENIX via YouTube

Tags

SREcon Courses Artificial Intelligence Courses Machine Learning Courses Deep Learning Courses Root Cause Analysis Courses Confirmation Bias Courses

Course Description

Overview

Explore a thought-provoking conference talk that challenges common misconceptions about Machine Learning (ML) in production engineering. Delve into why many proposed ML applications for Site Reliability Engineering (SRE) are structurally unsuitable for their intended purposes. Examine key problem domains where SREs aim to apply ML and understand why these applications often lack the necessary characteristics for feasibility. Learn to evaluate potential ML applications critically and discover approaches for determining their practicality. Gain insights into the limitations of ML in solving most desired problems while recognizing its potential for addressing specific issues. Through this 39-minute presentation by Todd Underwood from Google at SREcon19 Europe/Middle East/Africa, acquire a realistic perspective on ML's capabilities and limitations in the field of production engineering.

Syllabus

Intro
Agenda
Who am I
Machine Learning vs AI
Machine Learning Primer
Deep Learning vs ML
What is ML for
Why ML is bad
Ticket categorizing
Single ticket queue
Single ticket algorithm
Whats wrong with this
Automatic Root Cause Analysis
Outages
Statistical Correlation
Low Uncertainty
Deep Learning
Confirmation Bias
Plausible Use
Prework
ontology epistemology metaphysics
QA


Taught by

USENIX

Related Courses

How to Not Destroy Your Production Kubernetes Clusters
USENIX via YouTube
SRE and ML - Why It Matters
USENIX via YouTube
Knowledge and Power - A Sociotechnical Systems Discussion on the Future of SRE
USENIX via YouTube
Tracing Bare Metal with OpenTelemetry
USENIX via YouTube
Improving How We Observe Our Observability Data - Techniques for SREs
USENIX via YouTube