Explore the humorous side of machine learning failures in this 39-minute conference talk from Strange Loop. Discover how common algorithmic mistakes can lead to unexpectedly entertaining results, as presented by Janelle Shane from AIweirdness.com. Delve into examples of overfitting, noisy data, and overly general problems, learning how these issues can be deliberately used for creative purposes. Examine fascinating case studies, including cookie experiments, pun generation, neural net jokes, and image recognition mishaps. Gain insights into the challenges of knitting algorithms, adversarial attacks, and the importance of imperfection in AI development. Leave with a fresh perspective on embracing and learning from machine learning errors in both serious applications and artistic endeavors.
Overview
Syllabus
Introduction
Example of an experiment
Example of charming humans
Limited information
Experiment
Cookies
Surface appearances
Puns
Neural net jokes
April Fools jokes
Limited memories
Long memories
Unfair situation
Pony classification
Application image recognition
Fun to enjoy
Problem is too broad
Recipe for success
Another example
For example
It made some mistakes
Several in fact
Consistent
Clocks
Sheep
Wool
surrealism
sheep surrealism
spooky clocks
training data
boring pictures
pizza girl
image recognition algorithms
adversarial attacks
Skynet
Knitting example
Challenges
Knitting
Knitting without debugging
Tiny baby whales Soto
The knitters liked it
Perfection is always a bad thing
Take home message
Taught by
Strange Loop Conference
