# Generating Emotional Landscapes | Hannah Davis | OpenAI Scholars Demo Day 2018

## Метаданные

- **Канал:** OpenAI
- **YouTube:** https://www.youtube.com/watch?v=I61BnRBtUeQ
- **Дата:** 02.07.2020
- **Длительность:** 6:19
- **Просмотры:** 5,069
- **Источник:** https://ekstraktznaniy.ru/video/11597

## Описание

Hannah Davis talks about Generating Emotional Landscapes on OpenAI Scholars Demo Day on September 20, 2018. 

Learn more: https://openai.com/blog/openai-scholars-2018-final-projects#hannah

## Транскрипт

### Segment 1 (00:00 - 05:00) []

hi everyone I'm Hannah calling in from New York so I did my project not generating emotional landscapes and the background on this is that I've been pretty interested in just an emotional datasets for most of the past year and earlier in the year I created a emotion tag to landscape dataset where basically I took seven different classes of landscapes and used gram flour to have people tag them with eight different emotions and so some of the results I got from this included these aren't some emotion tagged fields so you can see that there's a pretty greater distinction between emotions here some disgust images mostly browns and greens a lot of water swampy Lake kind of things fear was like both obvious forests kind of dark blue nice area tech things but also like conceptual fear of the open ocean and surprise had a lot of like bright colors and things like that so I wanted to take this data set and see if I could actually generate emotional landscapes from it so first I tried again but that was very quickly unstable as you can see here and so I ended up trying a multi scale additional BAE which is based on Emily Denton's 2015 paper and so I'll show you some of these outputs these are 64 by 64 pixels generations so each of ocean row has the same defect there as the input so here's one set then I'll just go through them so here's anger anticipation disgust which you can see has a lot of grounds and greens fears pretty dark and gloomy Joy's a little brighter sadness is pretty muted surprise has some surprising colors I guess trust is very bright and calm here's another set and then I also tried doing landscape specific generations so here are 32 by 32 generated mountains slightly blurrier but I think the emotions still hold and then forests and that's my project great does anyone have any questions to directors now so the question was if people will agree to the emotions of the generated samples sorry do people agree on the generate examples if you show them a generated sample survey saying disgust or fear I think definitely if I show them like ten in a row oh it like like these kind of things where there's a sample of ten then they do there are some individual ones that don't totally feel right but as a whole they do I think like fear versus joy that looks pretty right to me any other questions I'm just going to bring hi so other than color colors were there other aspects of pictures that would correlate with emotions like symmetry of the image or certain shapes that appear in the image I actually noticed that anticipation had a lot of contrast more than I think more than most of the other ones someone said that maybe that was like a lens flare

### Segment 2 (05:00 - 06:00) [5:00]

effect or something so definitely like shadows and contrast between lighter and darker areas was another thing I had noticed but for the most part it did seem to be colors although with the mountains also the sky light seemed different so like fear the mountains all seem to be slightly higher than some of the other emotions which maybe makes sense if you're like looking up at a mountain versus like looking down from a mountain or something like that hey we've got another question when you were building the data set with CrowdFlower did you only select for English speakers or did you to translate those 10 emotions to different languages no I only selected for English speakers it's actually part of a bigger project I want to do which is like seeing if like people in different areas type things differently like if people who live near mountains type mountains differently and things like that so I got started in the US and like in you in people who I put in there US cities so I could like start investigating that