Deep Learning in Banking

List of Figures

1. 1.1 Example of a tabular dataset for supervised learning inspired by the credit scoring task.
2. 1.2 The architecture of the Perceptron model.
3. 1.3 Graphical representation of various activation functions.
4. 1.4 The architecture of the Multilayer Perceptron (MLP) model.
5. 1.5 Illustration of different data representations: scalar (0D), vector (1D), matrix (2D), and a three-dimensional tensor. While tensors can have any number of dimensions, the 0D, 1D, and 2D cases are so commonly used that they are given special names.
6. 2.1 Free LiDAR data coverage in the US Map services and data available from US Geological Survey, National Geospatial Program.
7. 2.2 Samples from the MNIST handwritten digits dataset.
8. 2.3 Example of a standard CNN.
9. 2.4 Functioning of the convolution and max pooling layers.
10. 2.5 Example of a LiDAR image for MSA 1018. Source: USGS data under open license.
11. 2.6 Loss plots for different learning rates. (a) Learning rate 10⁻³. MSE 0.0482 is higher than optimal. Training is too sudden, leading to some lower performance (local minima). (b) Learning rate 10⁻⁴. MSE 0.0452 is similar to optimal, but training is slightly sudden, leading to a tiny bit lower performance. (c) Learning rate 10⁻⁵. Optimal performance. MSE 0.045. The optimal value is reached at epoch 80, showcasing double descent. (d) Learning rate 10⁻⁶. MSE of 0.480 is higher than optimal, and training takes longer to converge. (e) Learning rate 10⁻⁷. MSE of 0.0556 is much higher than the optimal, meaning ...