[ireland352]

I was at a vehicle AI conference early this month discussing how AI will only output based on the quality of its inputs. And the fear coming out of that conference was if space wasn't correctly dedicated or allocated, the AI will optimize the systems as much as it can to squeeze out every single ounce of efficiency.

Qualitative items such as the ones listed are important for customer focused environments, however, I'm not sure if AI can account for such factors.

This post is quite timely.

[gwern]

> Qualitative items such as the ones listed are important for customer focused environments, however, I'm not sure if AI can account for such factors.

One of the nice things you can do with optimization problems is plug humans into the loop as oracles. Often, 'we know it when we see it', and we can do pairwise comparisons of 2 possibilities. So you can train a ML model based on win/loss comparisons and it'll learn to take into account the softer qualitative aspects via preference learning.

A recent example you might remember from the press: "Deep reinforcement learning from human preferences" https://arxiv.org/abs/1706.03741 , Christiano et al 2017 https://deepmind.com/blog/learning-through-human-feedback/ https://blog.openai.com/deep-reinforcement-learning-from-hum...

But also "Deep TAMER: Interactive Agent Shaping in High-Dimensional State Spaces" https://arxiv.org/abs/1709.10163 , Warnell et al 2017.

You can even pair it with EEG or brain scans for implicit ranking: "Towards personalized human AI interaction - adapting the behavior of AI agents using neural signatures of subjective interest" https://arxiv.org/abs/1709.04574 , Shih et al 2017.

[tw1010]

Like I mentioned, I have high hopes modelling vaguer aspects of customer preference can be incorporated by training a model on rendered generative design (e.g. with data from amazon turk), which can then be used as maybe some kind of penalization on the efficiency loss function, multiplied by a weight to give us a choice in the efficiency-comfort tradeoff-space.

There are solutions that I think show potential out there. I don't think our future AI designed world necessarily need to ignore difficult-to-quanitify dimensions like aesthetics. (Though amazon turk is expensive, especially in developer man-hours, so I can understand if that won't always be done.)

[lawlessone]

Could another parameter be how it crumples in a crash? I don't want my AI supercar to transform into inescapable cage on impact.

[ghein]

It all comes down to how explicitly preferences are originally understood and then if the reward function can incorporate implicit analysis.

There have been recent studies about AI powered shirt design - the original input uses existing designs in terms of color and shape rather than the basic naive description of requirements that an engineer would give. Then the designs can be assessed by a review board or put up on a site and not produced until some n quantity of purchases.

You wouldn't try to detect cats in images without labelled data why would you try something MUCH harder without labelled data?!?!?!?!?!