Computer-vision Classification-algorithms Are Inherently Creative When Error-prone

Abstract

Whether coming from a linear support vector machine, from logistic regression, or a quasi-Newtonian, the fine-tuning of the decision boundary in any given data set is essential to mitigate the loss term so that neural nets in image recognition can divide a data space into separate sections and correctly classify an input. By their very nature, neural nets are logically non-deterministic but rest on probability-weighted associations, which are adjusted recursively to enhance the similarity of intermediate results to the target output, the remaining difference being the ‘error.’ However, taxonomies should not be crisp but seen as fuzzy classes, allowing for hybrid exemplars that transgress category boundaries. The associative and similarity orientation of neural nets and deep learning makes such systems inherently creative in that misclassifications are at the basis of creative crossovers in information processing. This new conceptualization of network errors is supported by the ratings of 40 top-ranking designers of 20 image-recognition mistakes on the dimensions of creativity and innovativeness.