Skip to main content

Table 2 Chosen hyperparameters with brief description. Hyperparameter sweep went through a realizable range for each hyperparameter and individual values were chosen to optimize training ability or to minimize overfitting, depending on the parameter

From: Deep learning networks find unique mammographic differences in previous negative mammograms between interval and screen-detected cancers: a case-case study

Hyperparameter (Range) Hyperparameter Type Interpretation Chosen Value
Rotation (0–90) Data Augmentation Range for a random rotation 20
Zoom (0–1) Data Augmentation Range for a random zoom 0.5
Shear (0–1) Data Augmentation Range for a random shear 0.3
Vertical/Horizontal Flip (Yes/No) Data Augmentation Random chance of flip in respective direction Yes/Yes
Momentum (0–1) Optimizer Parameter Accelerates or dampens oscillations in given direction. 0.3
Regularization (0–1) Optimizer Parameter Penalty applied to large image weights 0
Decay (0–1) Optimizer Parameter Learning Rate decay over each update. 1e-5
Dropout (0–1) Fully-connected Layer Percent of weights dropped out between dense layers in the FC layer. 0.95
Learning Rate (0–1) Training Parameter Importance attributed to weight updates. 1e-3
Epochs (Integer) Training Parameter Number of epochs performed 1000
Batch Size (2n any n) Training Parameter Number of samples per gradient update 16
Image Size (Minimum 224) Training Parameter Input image size in pixels 224
nLayersRetrain (Fully Connected only – All Layers) Training Parameter Number of layers allowed to have their weights altered. All Layers (173)