This was fun to read. The website was fun to click. I agree that attention to that ineffable thing that might be called “consumer surplus” or “why we like things” is super interestingly important and understudied and related to a lot of “physical object appreciation” issues that are hard to write about.
I’d offer some abstract candidate things that people kinda like:
Bold clean contrasts (something is happening on purpose such that an error would be visibly disruptive).
Clicking though your website for labeling I found myself adopting an attitude of looking for something that it would make me actively happy to wear to work in an office, until I got bored of saying the same thing was great over and over and over.… and clicked on something I’d love to wear to a party. Then pretty fast I imagine I have to go back to work and slowly hill climbing through “preferable to wear to work vs a party dress (or the previous thing)” until I found a great work outfit again.
(As I did it, my data will exhibit circular preferences, I’m pretty sure. If there was a way to add a “salient as great” label, and change this when I change my mood or see a thing that calls to me and causes me to want a new label, then I think (or would like to imagine? or aspire to be such that...) my preference ordering under each label would be well ordered.)
I imagined you trying to train a visual model of “what Jennifer likes” (that wasn’t massively pre-trained to capture coherent articulate semantics from the images) and it didn’t seem likely to work… I don’t think I was picking things based on trivial 2D visual rules?
A lot of it was “oh I love the skirt and hate the top, but the outfit is better than my current best (like: I’d buy the outfit to keep the skirt)” and then “oh that top’s sleeves are fantastic, and the pants are tolerable” and so on.
I kind of expect MY labeling has a bunch of latent dimensions, and that everyone else’s clicking is also full of other dimensions (but also dimensions I care about) and that it would be really interesting to see the dimensional analysis, and a read out of my relative weights on the dimensions that your tool could identify.
If I have a guest ID, can I send it to you here and get outputs somehow, or do I have to start over my clicking and sign in if I want that?
Awesome, thank you for the thoughtful comment! The links are super interesting, reminds me of some of the research in empirical aesthetics I read forever ago.
On the topic of circular preferences: It turns out that the type of reward model I am training here handles non-transitive preferences in a “sensible” fashion. In particular, if you’re “non-circular on average” (i.e. you only make accidental “mistakes” in your rating) then the model averages that out. And if you consitently have a loopy utility function, then the reward model will map all the elements of the loop onto the same reward value.
Finally: Yes, totally, feel free to send me the guest ID either here of via DM!
Interesting! I’m fascinated by the idea of a way to figure out the transitive relations via a “non-circular on average” assumption and might go hunt down the code to see how it works. I think humans (and likely dogs and maybe pigeons) have preference learning stuff that helps them remember and abstract early choices and early outcomes somehow, to bootstrap into skilled choosers pretty fast, but I’ve never really thought about the algorithms that might do this. It feels like stumbling across a whole potential microfield of cognitive science that I’ve never heard of before that is potentially important to friendliness research!
This was fun to read. The website was fun to click. I agree that attention to that ineffable thing that might be called “consumer surplus” or “why we like things” is super interestingly important and understudied and related to a lot of “physical object appreciation” issues that are hard to write about.
I’d offer some abstract candidate things that people kinda like:
Bold clean contrasts (something is happening on purpose such that an error would be visibly disruptive).
Symmetry (because duh?).
Joyful colors (like in spring time).
Clicking though your website for labeling I found myself adopting an attitude of looking for something that it would make me actively happy to wear to work in an office, until I got bored of saying the same thing was great over and over and over.… and clicked on something I’d love to wear to a party. Then pretty fast I imagine I have to go back to work and slowly hill climbing through “preferable to wear to work vs a party dress (or the previous thing)” until I found a great work outfit again.
(As I did it, my data will exhibit circular preferences, I’m pretty sure. If there was a way to add a “salient as great” label, and change this when I change my mood or see a thing that calls to me and causes me to want a new label, then I think (or would like to imagine? or aspire to be such that...) my preference ordering under each label would be well ordered.)
I imagined you trying to train a visual model of “what Jennifer likes” (that wasn’t massively pre-trained to capture coherent articulate semantics from the images) and it didn’t seem likely to work… I don’t think I was picking things based on trivial 2D visual rules?
A lot of it was “oh I love the skirt and hate the top, but the outfit is better than my current best (like: I’d buy the outfit to keep the skirt)” and then “oh that top’s sleeves are fantastic, and the pants are tolerable” and so on.
I kind of expect MY labeling has a bunch of latent dimensions, and that everyone else’s clicking is also full of other dimensions (but also dimensions I care about) and that it would be really interesting to see the dimensional analysis, and a read out of my relative weights on the dimensions that your tool could identify.
If I have a guest ID, can I send it to you here and get outputs somehow, or do I have to start over my clicking and sign in if I want that?
Hi Jennifer!
Awesome, thank you for the thoughtful comment! The links are super interesting, reminds me of some of the research in empirical aesthetics I read forever ago.
On the topic of circular preferences: It turns out that the type of reward model I am training here handles non-transitive preferences in a “sensible” fashion. In particular, if you’re “non-circular on average” (i.e. you only make accidental “mistakes” in your rating) then the model averages that out. And if you consitently have a loopy utility function, then the reward model will map all the elements of the loop onto the same reward value.
Finally: Yes, totally, feel free to send me the guest ID either here of via DM!
Interesting! I’m fascinated by the idea of a way to figure out the transitive relations via a “non-circular on average” assumption and might go hunt down the code to see how it works. I think humans (and likely dogs and maybe pigeons) have preference learning stuff that helps them remember and abstract early choices and early outcomes somehow, to bootstrap into skilled choosers pretty fast, but I’ve never really thought about the algorithms that might do this. It feels like stumbling across a whole potential microfield of cognitive science that I’ve never heard of before that is potentially important to friendliness research!
(I have sent the DM. Thanks <3)