3 ms·

I think the biggest oversight is the actual feedback portion of sales to train the model on which features sold the best.

52-6F-62 9y ago
I was coming here looking for something like this -- a little more elaborate than this, but like this nonetheless. The algorithm could probably be nearly anything as long as it output a mix of complementary, or aesthetically pleasing colours and simple patterns. Psychology might do a better job designing labels than this algorithm. I understand the work of the algorithm is to produce the actual designs, so it accounts for the colours and patterns, but from a higher order all we are really seeing is: >> People like colour, make lots of colour. Simple is memorable, make simple. Punchy is eye-catching, make punchy. I would be more taken if there was indeed a feedback loop, or further testing on people to see exactly which patterns/colours had the highest number of takers, and a poll to corroborate whether or not they chose consciously or unconsciously.
- sogen 9y ago
  They could measure which designs sold faster: which pastel designs, vibrant colors, dots, shapes, etc., and get a quick overview of the preferences.
  - 52-6F-62 9y ago
    I think that's what would inform an interesting algorithm... but maybe I'm just conflating ideas. I think I'd be more smitten with a machine learning model that at first presented designs, then took as feedback sales data that informed its design progression. In the end, sure the patterns are catchy. I guess that's all people really care for in a Nutella label. :P
    - sogen 9y ago
      Yeah, for the second batch of jars take sales data and do A/B testing with a new round of designs!
  - lowglow 9y ago
    I agree there could definitely be a graduated approach to the granularity of features that might matter, but I think it would be possible to discover emergent multi-modal signals that matter.

52-6F-62 9y ago

I was coming here looking for something like this -- a little more elaborate than this, but like this nonetheless. The algorithm could probably be nearly anything as long as it output a mix of complementary, or aesthetically pleasing colours and simple patterns. Psychology might do a better job designing labels than this algorithm. I understand the work of the algorithm is to produce the actual designs, so it accounts for the colours and patterns, but from a higher order all we are really seeing is: >> People like colour, make lots of colour. Simple is memorable, make simple. Punchy is eye-catching, make punchy. I would be more taken if there was indeed a feedback loop, or further testing on people to see exactly which patterns/colours had the highest number of takers, and a poll to corroborate whether or not they chose consciously or unconsciously.

sogen 9y ago
They could measure which designs sold faster: which pastel designs, vibrant colors, dots, shapes, etc., and get a quick overview of the preferences.
- 52-6F-62 9y ago
  I think that's what would inform an interesting algorithm... but maybe I'm just conflating ideas. I think I'd be more smitten with a machine learning model that at first presented designs, then took as feedback sales data that informed its design progression. In the end, sure the patterns are catchy. I guess that's all people really care for in a Nutella label. :P
  - sogen 9y ago
    Yeah, for the second batch of jars take sales data and do A/B testing with a new round of designs!
- lowglow 9y ago
  I agree there could definitely be a graduated approach to the granularity of features that might matter, but I think it would be possible to discover emergent multi-modal signals that matter.