🧠 Dynamic Neural Network (Python From Scratch)

This project implements a fully customizable Feedforward Neural Network from scratch in Python — no external libraries like TensorFlow or PyTorch required. It allows you to define the number of layers, nodes, learning rate, and test it with new inputs.

---

📌 Features

• ✅ Automatically detects input/output layer size from training data
• ✅ User-configurable or random hidden layers
• ✅ Sigmoid activation function + derivative
• ✅ Forward pass & backpropagation
• ✅ Tracks best weights & biases based on lowest SSE (Sum of Squared Errors)
• ✅ Accepts new inputs after training for prediction

---

⚙️ How It Works

🔧 Model Setup

Input & Output nodes are detected automatically

3 modes for hidden layers:

Manual: User defines exact number of layers and nodes

Random: User specifies max limits; structure is randomized

Default: 5 hidden layers, 3 nodes each

🔁 Training Phase

Forward pass using sigmoid activation

Backpropagation using gradient descent

Weights & biases are updated per epoch

Model saves best parameters with minimum SSE

🔮 Prediction Phase

Predicts outputs for all training inputs

Accepts new inputs interactively and outputs predictions

---

🔢 Example Input Format

training_data = [ {"inputs": [0.2, 0.5, 0.9], "target": [1]}, {"inputs": [0.1, 0.3, 0.6], "target": [0]}, {"inputs": [0.7, 0.8, 0.2], "target": [1]}, ]

---

🧪 Sample Console Output

Choose one of the following options for hidden layers:

1. Predefined hidden layer count
2. Randomized layers and nodes
3. Default: 5 hidden layers, 3 nodes each

Enter your choice (1, 2, or 3): 3 Enter the Learning Rate: 0.1 Enter number of epochs: 1000

Training completed! Least SSE: 0.002194

Predicted outputs for trained inputs: Input: [0.2, 0.5, 0.9] -> Predicted Output: [0.9712]