The problem we want to address
AI(Artificial intelligence) is diving into people’s lives as its algorithm continues to iterate. However, the algorithmic and quantitative systems do not seem to access people’s experiences, which always include emotional and qualitative factors. How can an AI system understand people’s feelings? How can an AI system like GANs optimize for meaningful human values, like beauty?

Our approach
In this project, I mainly build two approaches for designers to translate the human experience into GANs. Starting from the exploration of GANs, I investigate their potential chances to “understand” human experience. Then, the qualitative research
on human experience reveals the aesthetic factors influencing people’s feelings
about GANs‘ output. Through the interviews, some pain points from AI product
designers are concluded. Combining the exploration results in those three scopes, I
hypothesize that retraining with the highly-rated images and building new computational models to iterate the factors from the human evaluation can help designers inform the AI system about the human experience. I cooperated with CSE students and proposed two approaches- CURATION APPROACH and ALGORITHMIC
AESTHETIC APPROACH- based on the above assumptions. The build process demonstrates the achievability of the assumptions. In the evaluation part, the retrained images in both methods are rated higher than the original ones. Both approaches are proved to be practical and feasible.

Our result
For the curation approach, all produced models with highly-rated images outperformed the original model. The original model’s score is 0.2. When replacing 500,1000,2985 highly-rated images into the input dataset, their scores increased to 0.245, 0.269, and 0.255. For the algorithmic aesthetic approach, we select three categories of images(coastline, forest/desert, arctic). In each category of images, three factors are selected to improve. Among the nine groups whose corresponding factors are iterated by algorithms, eight groups’ new images are rated higher by people than the original images. The curation and algorithmic aesthetic approaches are verified to help inform AI systems of human experience.

Potential impact
Following the curation and algorithmic approaches, designers can successfully inform GANs about people’s aesthetic evaluation of their images. Using the human experience to improve AI systems is a starting point. It also proves the importance of the human experience for AI and provides a template for designers in all AI fields to inform their AI systems of human experience.