Search Papers | Poster Sessions | All Posters
Poster B79 in Poster Session B - Thursday, August 8, 2024, 1:30 – 3:30 pm, Johnson Ice Rink
The Role of Image Quality in Shaping Neural Network Representations and Performance
Jason Lee1, Prince Owusu Nkrumah1, Stephen Chong Zhao1, William J. Quackenbush1, Adaline Leong1, Trisha Mazumdar1, Mark Wallace1, David A. Tovar1; 1Vanderbilt University
Neural networks, recognized as robust models of the brain, depend on various factors including architectures, training data, algorithms, and objective functions. This study explores the influence of image quality in training data on the representation and performance of neural networks, and consequently, on their capability to model brain functions. The "visual diet"—the quality and variety of images—present in training sets such as ImageNet and EcoSet, can significantly affect how these models learn and perform across different classes. By examining how variations in image quality impact the networks' internal representations and overall performance, we aim to better understand how training data affects the correspondence between neural network models and the brain's processing mechanisms. Our main finding is that high-quality training data is one which encompasses a diverse range of image quality metrics, with our findings indicating that image categories with image quality diversity exhibit the most expansive representational spaces and the highest performance.
Keywords: Image Quality Representations Brain Models Neural Networks