Fig. 7: Investigation of the feature space of ResNet50 network applied on a public COVID-19 dataset for classification into four categories.
From: Revealing hidden patterns in deep neural network feature space continuum via manifold learning
a–c t-SNE, UMAP, and MDA visualizations of the feature spaces at four different layers before/after training. Here, S2-B4-L3 denotes the 4th residual block’s last convolutional layer in substructure 2, S3-B2-L3 denotes the 2nd residual block’s last convolutional layer in substructure 3, S3-B6-L3 denotes the 6th residual block’s last convolutional layer in substructure 3, and S4-B3-L3 denotes the 3rd residual block’s last convolutional layer in substructure 4. Before training, the data points are randomly distributed in MDA visualizations. However, after the training, the feature space becomes well clustered in MDA visualizations, especially in deeper layers. t-SNE and UMAP fail to show any information about the training status of the network. d k-nearest neighbor classification accuracy of the low dimensional representations from different techniques. Source data are provided as a Source Data file.
