AlexNet hallucination quiz: can you find which image matches the stated class?

robots.ox.ac.uk

28 points·stared··9 comments

Some context on image classifier trained on ImageNet dataset here: https://github.com/stared/thinking-in-tensors-writing-in-pyt... (a Jupyter/Colab notebook, so you can use various networks, including AlexNet).

Such hallucinations are related to https://distill.pub/2017/feature-visualization/ (also with Colab code how to create such).

Who is the intended audience? Some of them appeared to be westernized Japanese words, so even after seeing which ones were the right answer, I still don't know what those ones were.

Other than that, it is kinda cool searching through the distortions.

The most likely audience are probably machine learning researchers, because they'd know about the ImageNet ILSVRC classes: http://image-net.org/challenges/LSVRC/2014/browse-synsets

If you see what appears to be transliterated Japanese and it's not "kimono", you're probably dealing with some kind of animal, like an Agama https://en.wikipedia.org/wiki/Agama_(lizard) or a Basenji https://en.wikipedia.org/wiki/Basenji , neither of which are Japanese.

It's a reverse captcha. Only robots can answer all images with ease, while humans will have difficulties.

I didn't have any problem at all. Couldn't find one I didnt get on the first try. I fail maybe 30% of captchas haha.

U, Robot.

I was 9 for 9, then I got asked to select a "palace". I didn't really know how to visualize one, but there was a picture reminiscent of a an clearly very ornate chair in a room, which I figured could be a palace interior. I broke my streak choosing that chair :(

Got most of them right when I allowed myself search for pictures of the word to know what I was looking for.

This is actually a really fun game

Reply on news.ycombinator.com