ZERO DEPENDENCY (DENSE MLP) NEURAL NETWORK

Key Architectural Features

• Zero-Dependency Core: Numpy is fully used as the main mathematical source.
• Numerical Stability Engine: Handles forward and backpropagation like a champ. Has following modes:
  • Sigmoid : Has Overflow clipping, anti-vaneshing mechanism with hybrid RelU style gradients.
  • RelU : Has fixed dead neuron problems with Leaky RelU.
• Explicit Gradient-Wise Updates: Hand-rolled backpropagation executing exact partial derivatives ($\frac{\partial L}{\partial W}$, $\frac{\partial L}{\partial b}$) to optimize weights layer-by-layer.
• Modular Pipeline Design: Highly flexible design with various pipeline such as:
  • Image Pipeline : Auto cropping of handwritten numbers using cv2 (format: --> Non_Inverse)
  • Zip Pipeline : Handles zipped folders .zip to extract required data into specific format for training samples.
    • Image extraction (MINST Data)
    • Prediction data (Titanic Data sets and e.t.c)
• Handwritten Mini-Batch Training: Small amount data are calculated normally but anything that exceeds 32 examples are taken mini batch training process.