Auflistung nach Autor:in "Linse, Christoph"

Gerade angezeigt 1 - 1 von 1

Towards understanding convolutional neural networks through visualization and systematic simplification
(2025) Linse, Christoph
Black-box systems like Convolutional Neural Networks (CNNs) have transformed the field of computer vision. While visualization tools have helped explore and explain CNNs, their inner workings remain opaque, particularly how they detect specific features. As deep learning applications become more widespread across various fields, it becomes crucial to understand these models. This understanding is needed to avoid misinterpretation and bias, which can seriously affect society. This research motivates holistic visualization approaches, which show various aspects of CNNs. Existing visualizations often focus on a few aspects, answering specific questions. Combining them in comprehensive software could provide a more holistic view of CNNs and their inner processes. While 2D space cannot present all relevant information due to screen size restrictions, 3D environments offer new representation and interaction opportunities. Therefore, we enable the visualization of large CNNs in a virtual 3D space. This work further contributes to the visualization field by improving the activation maximization method for feature visualization, which previously struggled with local maxima. In addition to visualization, this research increases CNN transparency through systematic simplification. We use pre-defined convolution filters from traditional image processing in modern CNN architectures. Instead of changing the filters during training, the training process finds linear combinations of the pre-defined filter outputs. Our Pre-defined Filter Convolutional Neural Networks (PFCNNs) with nine distinct edge and line detectors generalize better than standard CNNs, especially on smaller datasets. For ResNet18, we observed increased test accuracies ranging from 5-11 percentage points with the same number of trainable parameters across the Fine-Grained Visual Classification of Aircraft, StanfordCars, Caltech-UCSD Birds-200-2011, and the 102 Category Flower dataset. The results imply that many image recognition problems do not require training the convolution kernels. For practical use, PFCNNs can even save trainable weights.