What is it about?
Neural networks excel at analyzing images for many purposes, from finding cats to judging the authenticity of a depicted artwork. But neural networks are data hungry; training them typically requires thousands, even millions, of examples. While there may be that many cats, there are only a few hundred Rembrandts. How can we train a neural network to make fine distinctions among artists based on a small oeuvre of works?
Featured Image
Photo by redcharlie | @redcharlie1 on Unsplash
Why is it important?
We describe an approach that decomposes source images into many small, overlapping tiles, and then sifts them for information diversity: the greater the complexity of the tile content, the more effective it will be as a training example. Using this approach allows us to train neural networks that distinguish the paintings of Rembrandt and van Gogh from their contemporaries, admirers, and forgers, despite the relative dearth of available works.
Read the Original
This page is a summary of: Salient Slices: Improved Neural Network Training and Performance with Image Entropy, Neural Computation, June 2020, The MIT Press,
DOI: 10.1162/neco_a_01282.
You can read the full text:
Contributors
The following have contributed to this page
