• Mathematical background: "Neural Networks and Deep Learning".

• Datasets: UCI Machine Learning Repository.

• The code has been written and tested in Python 3.7.7.
• Multi-input/multi-output (multivariate) feed-forward neural network implementation for regression and classification.
• Arbitrary number of nodes for input, hidden, and output layers.
• Continuous problem: quadratic cost function, L2-type regularization term, sigmoid activation for hidden layers, linear activation for output layer.
• Classification problem: cross-entropy cost function, L2-type regularization term, sigmoid activation for hidden and output layers, classes determined automatically.
• Sigmoid and cross-entropy function are computed using a numerically stable implementation.
• Gradient of the cost function calculated using the backpropagation algorithm.
• Option to reduce the learning rate during the computation.
• Option to not to compute and return the gradient.
• A gradient descent optimizer (GDO) is included in utils.py, together with several other utility functions.
• The FFNN class in FFNN.py is not constrained to the GDO solver but it can be used with any other optimizer.
• Usage: python test.py example.

example Name of the example to run (wine, stock, wifi, pulsar.)

problem Defines the type of problem. Equal to C specifies a classification problem, anything else specifies a continuous problem. The default value is None.

use_grad Specifies if the gradient is calculated and returned. The default value is True.

init_weights Specifies if the weights are randomly initialized. The default value is True.

data_file File name with the dataset (csv format).

n_features Number of features in the dataset (needed only for continuous problems).

hidden_layers List, tuple, or array with the number of nodes in each hidden layers.

split_factor Split value between training and test data.

L2 Regularization factor.

epochs Max. number of iterations (GDO).

alpha Learning rate (GDO).

d_alpha Rate decay of the learning rate (GDO).

tolX, tolF Gradient absolute tolerance and function relative tolerance (GDO). If both are specified the GDO will exit if either is satisfied. If both are not specified the GDO will exit when the max. number of iterations is reached.

There are four examples in test.py: wine, stock, wifi, pulsar. Since GDO is used, use_grad is set to True. For all examples init_weights is also set to True.

data_file = 'wine_dataset.csv'
n_features = 11
hidden_layers = [20]
split_factor = 0.7
L2 = 0.0
epochs = 50000
alpha = 0.2
d_alpha = 1.0
tolX = 1.e-7
tolF = 1.e-7

Original dataset: https://archive.ics.uci.edu/ml/datasets/Wine+Quality.

The dataset has 11 features, 1 label, and 4898 samples.

The neural network has a layout of [11, 20, 1] and 261 variables.

Predicted/actual correlation values: 0.708 (training), 0.601 (test).

Exit on epochs with tolX = 2.0e-4 and tolF = 1.1e-7.

data_file = 'stock_dataset.csv'
n_features = 6
hidden_layers = [4, 4]
split_factor = 0.70
L2 = 0.0
epochs = 50000
alpha = 0.99
d_alpha = 1.0
tolX = 1.e-7
tolF = 1.e-15

Original dataset: https://archive.ics.uci.edu/ml/datasets/ISTANBUL+STOCK+EXCHANGE.

The dataset has 6 features, 3 labels, and 536 samples.

The neural network has a layout of [6, 4, 4, 3] and 63 variables.

Predicted/actual correlation values: 0.841 (training), 0.840 (test).

Exit on epochs with tolX = 4.7e-6 and tolF = 9.8e-11.

data_file = 'wifi_dataset.csv'
problem = 'C'
hidden_layers = [10, 5, 4]
split_factor = 0.70
L2 = 0.0
epochs = 50000
alpha = 0.9
d_alpha = 1.0
tolX = 1.e-7
tolF = 1.e-10

Original dataset: https://archive.ics.uci.edu/ml/datasets/Wireless+Indoor+Localization.

The dataset has 7 features, 4 classes, and 2000 samples.

The neural network has a layout of [7, 10, 5, 4, 4] and 179 variables.

Predicted/actual accuracy values: 100.0% (training), 98.0% (test).

Exit on epochs with tolX = 3.9e-5 and tolF = 1.0e-8.

data_file = 'pulsar_dataset.csv'
problem = 'C'
hidden_layers = [10, 10]
split_factor = 0.70
L2 = 0.0
epochs = 5000
alpha = 0.9
d_alpha = 1.0
tolX = 1.e-7
tolF = 1.e-7

Original dataset: https://archive.ics.uci.edu/ml/datasets/HTRU2.

The dataset has 8 features, 2 classes, and 17898 samples.

The neural network has a layout of [8, 10, 10, 2] and 222 variables.

Predicted/actual accuracy values: 98.1% (training), 98.0% (test).

Exit on epochs with tolX = 2.5e-4 and tolF = 5.5e-7.