Final Report Summary - CREATIV (Creating Co-Adaptive Human-Computer Partnerships)
The ERC CREATIV project worked closely with creative professionals, including designers, composers, choreographers, and scientists, to explore a radically different approach to what we call 'human-computer partnerships'. Instead of assuming that the goal of the computer is to classify human behavior to achieve an optimal answer, and discarding human variation as 'just noise', we reveal aspects of the system's behavior to the user, in a form they can understand. Our goal is to foster human creativity while harnessing the power of the machine –– rather than deskilling people and replacing them with a machine, we augment human capabilities and leave the user in control.
Our approach challenges traditional 'human-in-the-loop' machine learning research, which measures success in terms of improved algorithms rather than enhanced human ability or creativity. When humans perform an action, they generate a complex trace that includes both unconscious and intentional individual variation, as well as variation based on current external factors and random noise. For example, we can easily identify human characteristics in handwritten text, and guess not only who wrote it, but under what circumstances: carefully, rushed, or perhaps, on a bus. Instead of ignoring this variation, we examine it and display it progressively, allowing users to both understand what the system is interpreting and control it. Revealing the characteristics of human input to a machine learning algorithm opens up new integrated feedback loops that leave users in control, with as much or as little assistance as they desire. Users remain in control of the interaction, with an incrementally learnable path from novice to expert. Unlike systems that seek to solve a specified problem for the user, we demonstrate more creative activities where people and computers share agency, each taking advantage of the other's unique skills.
A key contribution of the CREATIV project is what we call 'generative theory', which provides an underlying 'physics of information' as a foundation for helping users predict and control their interactions, while offering inspiration for novel tools that retain simplicity of interaction, with significantly greater power of expression. Rather than replacing users by performing tasks for them, such systems augment human capabilities. Here, the combination of human and computer is greater than what is possible with either alone.
Our approach challenges traditional 'human-in-the-loop' machine learning research, which measures success in terms of improved algorithms rather than enhanced human ability or creativity. When humans perform an action, they generate a complex trace that includes both unconscious and intentional individual variation, as well as variation based on current external factors and random noise. For example, we can easily identify human characteristics in handwritten text, and guess not only who wrote it, but under what circumstances: carefully, rushed, or perhaps, on a bus. Instead of ignoring this variation, we examine it and display it progressively, allowing users to both understand what the system is interpreting and control it. Revealing the characteristics of human input to a machine learning algorithm opens up new integrated feedback loops that leave users in control, with as much or as little assistance as they desire. Users remain in control of the interaction, with an incrementally learnable path from novice to expert. Unlike systems that seek to solve a specified problem for the user, we demonstrate more creative activities where people and computers share agency, each taking advantage of the other's unique skills.
A key contribution of the CREATIV project is what we call 'generative theory', which provides an underlying 'physics of information' as a foundation for helping users predict and control their interactions, while offering inspiration for novel tools that retain simplicity of interaction, with significantly greater power of expression. Rather than replacing users by performing tasks for them, such systems augment human capabilities. Here, the combination of human and computer is greater than what is possible with either alone.