Open Access
Table B.1.
NN layers and their role in the model architecture.
| Layer Type | Acronym | Description |
|---|---|---|
| 1D Convolution | Conv1D(n, k) | Applies n filters of size k across 1D input sequences (e.g. photometric spectra), enabling local pattern extraction such as flux variations between adjacent bands. |
| 2D Convolution | Conv2D(n, k × k) | Applies n 2D filters over an image to extract spatial features. In this context, it helps capture galaxy morphology including compactness, shape, elongation, and brightness gradients. |
| Max Pooling (1D) | MaxPooling1D(p) | Reduces the dimensionality of 1D data by selecting the maximum value over a pooling window of size p. Helps retain prominent spectral features while reducing computational cost. |
| Max Pooling (2D) | MaxPooling2D(p × p) | Downsamples 2D feature maps by selecting the maximum value within each p × p window. Reduces spatial resolution while preserving key visual features. |
| Flatten | – | Converts multi-dimensional data (e.g. tensors) into a 1D vector suitable for input to fully connected layers. |
| Fully Connected | Dense(n) | A dense layer with n neurons that combines and transforms input features using learned weights. Used for non-linear integration of extracted features. |
| Dropout | Dropout(p) | Randomly disables a fraction p of neurons during training to prevent overfitting and encourage generalization. |
| Concatenation | – | Merges multiple input vectors into a single unified feature representation, used here to combine the outputs of the three branches. |
Current usage metrics show cumulative count of Article Views (full-text article views including HTML views, PDF and ePub downloads, according to the available data) and Abstracts Views on Vision4Press platform.
Data correspond to usage on the plateform after 2015. The current usage metrics is available 48-96 hours after online publication and is updated daily on week days.
Initial download of the metrics may take a while.