Université du Québec à Montréal, Montreal, Quebec, Canada
November 18-21, 2017
The "Painter By Numbers" Kaggle competition challenged participants to come up with an algorithm which can examine a pair of paintings and make a prediction: are these two paintings by the same artist? This talk explores some of the technical challenges involved with making an exciting and challenging competition, including how to build a simple RandomForestClassifier algorithm with sklearn as well as how to reduce the problem space to a manageable size by using painting metadata and clustering algorithms to identify groups of similar paintings.
I also explain how to implement siamese neural networks in keras and examine how they allow the development of classification algorithms that can learn to extrapolate to new classes. Finally, I explore how the winning algorithm performed in an actual case of art fraud.
Kiri is a lapsed physicist who honed her machine learning chops on Kaggle where she has attained Master status. When she's not Kaggling, Kiri does consulting work helping companies to prototype machine learning tools.