distillpub / post--feature-visualization

Feature Visualization
https://distill.pub/2017/feature-visualization/
Creative Commons Attribution 4.0 International
128 stars 33 forks source link

Just some small typos #14

Closed rdipietro closed 6 years ago

rdipietro commented 6 years ago

Visualizations of all channel are available in the appendix. (channels)

This article focusses on feature visualization. (focuses)

In the following sections we’ll examine techniques to get diverse visualizations, understand how neurons interact, and avoid high frequency artefacts. (artifacts)

How can we chose a preconditioner that will give us these benefits? (choose)

used in adverserial work (adversarial)

As long as a correlation is consistent across spatial positions — such as the correlation between a pixel and its left neighbor being the same across all positions of an image — the Fourier coefficients will be independant variables. (independent)

Canas commented 6 years ago

I'll add a few others I found :)

The objectives we’ve mentioned only scratch the surface of possibile objectives. (possible)

Additionally, the optimisation may make examples be different in an unnatural way. (optimization, for consistency)

Gradient blurring[15] is equivalent to gradient descent in a different paramaterization of image space, (parameterization)

ludwigschubert commented 6 years ago

Yay for all your work and attention! Will fix on Monday. :-)