Neural Network
Implementation of a feedforward (fully connected) Neural Network from scratch
in C++, based on the object-oriented programming paradigm. Accordingly, classes
have been created in increasing order of abstraction to define the topology of
this neural model: Neuron.hh
, Layer.hh
and NeuralNetwork.hh
. Keeping in
mind the ease of use for the user and allowing the greatest possible
flexibility, classes are also designed for reading data (Data.hh
) and for
configuring the network (Custom.hh
).
Getting Started 🚀
These instructions will get you a copy of the project up and running on your local machine for development and testing purposes.
Installing 🔧
Clone the repository.
git clone https://gitlab.com/david-alvarez-rosa/neural-network.git
Compile the code (it may take some time the first time).
cd neural-network/
make
Running the Algorithm ️
Once the code is compiled just execute main
,
./main
and enjoy.
Note. The user’s configuration together with the reading of the data must have been modified in the corresponding files prior to its execution.
Understanding the Code 💡
Paper with implementation coming soon.
Objectives 🎯
different objectives were had in mind when carrying out this project. They stand out among them:
- Create a modular codebase using object-oriented programming.
- Perform the gradient computation (i.e., partial derivatives of the loss function) automatically, accurately (not numerically) and efficiently. The algorithm used here is a recursive algorithm proposed by the author, as opposed to the one known as backpropagation.
- Create code that is easy to use by a user with no prior knowledge of Machine Learning.
- Allow the user as much flexibility as possible: from the choice of activation functions (including neuron to neuron) to the number of layers (with their corresponding number of neurons), through loss functions.
Difficulties and Future Work 🛠
The most complex part of the training algorithm of a neural network is the computation of the gradient of the loss function with respect to the model parameters (i.e., the computation of each of the partial derivatives). Here we propose a recursive algorithm adapted exclusively to the topology of this type of networks, making extensive use of the chain rule for the computation of derivatives of composite functions (in its multivariate version). However, thinking about generalization to other more complex models (e.g., convolutional neural networks or recurrent neural networks), it would be more interesting to use graphs to represent the models, and, in this way, to be able to make use of the chain rule iterating through the nodes of the graph in reverse order.