Feedforward Neural Network from Scratch

Project roadmap

1. 01

   Implement matrix and vector primitives

   ~2h

   Write functions for matrix multiplication, addition, and element-wise activation application using flat arrays.

2. 02

   Build the network structure

   ~1.5h

   Define layers with weight matrices and bias vectors, initialized with small random values.

3. 03

   Implement forward propagation

   ~2h

   Compute layer outputs by applying weights, biases, and a sigmoid/ReLU activation function through the network.

4. 04

   Implement backpropagation and gradient descent

   ~3.5h

   Compute output error, propagate gradients backward through layers, and update weights with a learning rate.

5. 05

   Load a small training dataset

   ~1.5h

   Parse a simple CSV dataset (or generate synthetic data) into input/label arrays for training.

6. 06

   Train and evaluate the network

   ~2h

   Run training for multiple epochs, track loss, and evaluate accuracy on a held-out test split.

Resources

• DocsC Language Reference