Human-in-Loop Computational Creativity

随着深度学习的最新进展，计算创造力已经成为一个突出的研究领域。目前的系统包括风格转移网络、创意素描生成[1]和肖像艺术[2][3][4]，可以分类为人类作为策展人、人类策展人作为数据源或人类作为合作者[5]。这些系统缺乏来自人类的一定程度的创造性控制，因为要么人类提供输入（草图的起跑线），然后系统完成草图，要么系统有输出，而人类充当管理者。在这两种情况下，人类都无法对过程进行控制，而是被限制在过程的开始或结束时进行控制。为了解决这个问题，提出了一个人在环（HIL）系统，允许人类根据自己的创造力修改或约束系统的输出。人脸是生成式人工智能系统和计算创造力的共同关注点。其原因是它们具有高度的规律性。漫画也有高度的规律性，稍后将在提案中探讨。结合其结构简单和备用的性质，它们是一种极好的媒介，通过它可以探索HIL网络在计算创造力方面的应用。由于卡通可以作为一系列相连的部分生成，它也提供了一个起点，以解决通过独特的部分组合从学习分布生成真正的新事物的问题。

With recent advances in deep learning computational creativity has come to prominence as an area of study. Current systems ranging from style transfer networks, creative sketch generation [1] and portrait art [2][3][4] can be categorized as either human as a curator, human curatorship as data source or human as a collaborator [5]. These systems lack a degree of creative control from the human as either the human provides an input (a starting line for a sketch) and the system completes the sketch, or the system has an output and the human acts as a curator. In both cases the human is unable to exercise control over the process, instead being limited to having control at the beginning or end of it. To solve this a Human-In-Loop (HIL) system is proposed that would allow for a human to modify or constrain the outputs of the system based on their creativity. Human faces are a common focus of generative AI systems and computational creativity. The reason for this is that they have high degree of regularity. Cartoons also have high degree of regularity as will be explored later in the proposal. Combined with their structurally simple and spare nature they are an excellent medium through which the application of HIL networks to computational creativity can be explored. As cartoons can generated as a series of connected parts it also provides a starting point to solve the problem of generating something authentically new from a learned distribution through unique combinations of parts.