A learning project where we attempt to build a Fully-Connected neural network from stratch. The purpose here is clearly not to provide a performant model, but rather to learn the inner mecanisms (backprop, forward pass, etc.) of high-level libraries that abstracts those from the user.
A Fully-Connected neural network with a variable number of hidden layers and variable number of nodes per layer. The model is applied on the well known fashion-mnist dataset.
Our tests are being done with 5 hidden layers, having 300 nodes each.
numpy
keras
matplotlib.pyplot
time
Approximately 4h00 were required to train and test the model, on a standard laptop without any use of the GPU (no parallelization has been implemented), and for only 50 epochs : clearly the model is unefficient.
With untuned hyperparameters we obtain a final precision on the validation set of approximately 88%.
Edward H-Hannan