I wanted to see how these algorithms actually work under the hood. No TensorFlow, no PyTorch, no "magic" libraries—just raw NumPy and the math I learned from diving deep into linear algebra and calculus.

Most ML tutorials just tell you to import a library and call .fit(), which feels like cheating. This project is my attempt to build the "engines" from scratch: implementing the gradients, the backprop, and the decision logic from the ground up.

I built a cool interactive script to show everything off. You can pick an algorithm, watch it train, and see the graphs pop up in real-time.

# Start the interactive menu
python run_all.py

If you're in a hurry and just want to run one thing from the terminal:

python run_all.py linear cnn # runs specific ones
python run_all.py rnn # runs the Hamlet text generator

I wrote down the math and the "why" for each algorithm. If you want to see the derivatives or how the convolution stride tricks work, check these out:

• 1. Linear Regression (The starting point: fitting a line)
• 2. Logistic Regression (Stepping up to classification)
• 3. K-Nearest Neighbors (Geometric similarity)
• 4. Gaussian Naive Bayes (Probabilistic classification)
• 5. The Bridge: Neural Networks (Combining it all into a modular MLP)
• 6. Convolutional Neural Network (The spatial challenge for images)
• 7. Recurrent Neural Network (The "final boss" of bookkeeping and memory)

I spent a lot of time on the plots because seeing the math is way better than just looking at loss numbers:

• 3D Loss Bowls: Watch the gradient descent fall down the convex surface.
• Saliency Maps: See exactly which pixels the CNN thinks are important.
• Hidden State Heatmaps: Peek into how the RNN's "brain" remembers characters.

.
├── documentation/ # Where I explain all the math
├── plots/ # All the generated graphs
├── run_all.py # The main entry point
└── [Algos]_From_Scratch/ # The actual source code