This model is a handmade neural network (without libraries such as TensorFlow or PyTorch) that has been trained to add 2 numbers.
- architecture: 2 inputs, 5 hidden neurons (ReLU), 1 output neuron.
- Dataset: 1.000 unike sums between 0 and 10.
- Trainingduration: 1.000.000 epochs.
- Total calculations: 1 billion sums (1.000.000 epochs × 1.000 samples).
- Learning Rate: 0.001
The model performs perfectly within the range 0-20.
When extrapolating (numbers above 100), a small linear deviation of approximately 2% occurs because the learned slope (weight) is approximately 0.98 instead of exactly 1.00.