BasicDNN

Historical learning project from 2020-2021. This repository is retained as an educational artefact and does not represent production AI/ML engineering.

Status and Limitations

• Not actively maintained.
• No automated test suite, packaging, CI or production deployment.
• Some advanced training paths may be incomplete or incorrect.
• Not intended for production use.

Generic L-layer fully connected neural-network implementation written directly in Python using NumPy.

Implemented Concepts

• Input data arranged as an n_x by m matrix, where n_x is the number of input features and m is the number of training examples.
• Output data arranged as a 1 by m matrix.
• Activation functions: Sigmoid, ReLU, Leaky ReLU, Tanh and Softmax.
• Weight initialization: zeros, random, He and Xavier.
• Regularization: L2 and Dropout.
• Optimization: Mini-Batch Gradient Descent, Momentum and Adam.

Historical Usage Example

MODEL = (
    (20, "relu", "he"),
    (7, "relu", "he"),
    (5, "relu", "he"),
    (1, "sigmoid", "random"),
)

parameters = L_layer_model(trainX, trainY, MODEL)
predictTrain = predict(trainX, parameters, trainY)
predictDev = predict(devX, parameters, devY)
predictTest = predict(testX, parameters, testY)

Batch normalization was planned but not implemented.