Creating a Human Similarity Ratings Benchmark Database for Artificial Neural Networks
Name
Taavi Gilden
Abstract
In recent years, there has been a growing interest in understanding the way in which humans encode and categorise visual images. One hope is that this knowledge can be translated into improvements in artificial agents that carry out image classification automatically. One approach for closing in on the code of human object categorisation is to understand the relational coding of different objects in terms of their similarity. However, due to the large number of possible objects and their even larger number of possible comparisons, classical experimental approaches are limited in building representational similarity spaces.
The goal of this project was to create a web-application, which could be accessed by a large number of people around the world, who could contribute to the goal of the project. The webpage allows subjects to rate natural images according to their perceived similarity. This yields a large and reliable dataset of image ratings that could be accessed and used by experimenters, but also by image classification experts worldwide. Based on the dataset a similarity matrix can be constructed. The webpage and the corresponding database will hopefully advance our understanding of human vision and may one day be used to improve artificial intelligence algorithms designed to categorise objects.
Graduation Thesis language
English
Graduation Thesis type
Bachelor - Computer Science
Supervisor(s)
Jaan Aru, Raul Vicente, Martin Hebart
Defence year
2015