Visualizing Discriminative Features with t-SNE: A Comparative Study of Model Architectures

t-SNE has been widely used to visually represent high-dimensional data in a low-dimensional space using nonlinear techniques. In our study, we utilize t-SNE to demonstrate the visualization results of linear features on the final layer of our model. We apply feature visualization on both the backbone model and the proposed model. To improve the visualization, we randomly select 30 categories from the CUB-200-2011 dataset for our experiment.

From Fig.6 to Fig.8, it is evident that our proposed method learns more distinctive visual representations. In Fig.6, the benchmark network TinyVit is able to group most similar samples together, but there are still a few samples that are grouped into distant regions. In Fig.7, the feature with only the SFRL module added is generally able to group similar samples into a single region, but there are some regions where categories are cross-confused. In Fig.8, we combine the SRFL and OL modules. After the low-dimensional processing, the high-dimensional features output by the overall model significantly reduce the occurrence of cross-regions between different categories.