aino 1.0.2

AINO Is Neural Operation: A lightweight, educational neural network library.

AINO (Aino is Neural Operation)

"Aino is Neural Operation." > A custom-built Deep Learning framework built from scratch using pure Python and NumPy.

---

🌟 Inspiration

This project was born out of curiosity after watching this inspiring video:
MIT Introduction to Deep Learning | 6.S191

I didn't want to just "import tensorflow" and call it a day. I wanted to see directly how the magic happens under the hood. I wanted to feel the weight of the matrices, understand the flow of the gradients, and build the brain neuron by neuron.

The Tech Stack

So, I built AINO using only NumPy.

• No Black Boxes: Every Forward Pass and Backpropagation step is manually calculated.
• Modular Design: Built with Object-Oriented Programming (Perceptron -> Layer -> NeuralNetwork).
• Custom File Format: Saves and loads trained models using the custom .dit format.

Features

• Flexible Architecture: Define any number of layers and neurons (e.g., [3, 20, 10, 1]).
• Activation Functions: Supports Sigmoid, ReLU, Tanh, and more soon.
• Tasks: Capable of both Classification (e.g., Iris Dataset) and Regression (e.g., Rocket Trajectory Prediction).
• Training Engine: Includes .fit() and .predict() methods similar to Scikit-Learn.

🧠 What I Learned

Building AINO from the ground up gave me insights that high-level libraries often hide:

1. Object-Oriented Neural Architecture: Instead of abstract tensors, I built the network hierarchically. I learned how a Neural Network manages Layers, and how a Layer orchestrates individual Perceptrons. This OOP approach helped me visualize exactly how data flows through the structure.

2. The Calculus of Backpropagation: I implemented the Chain Rule manually. By coding the backward() method in Layer and update_weight() in Perceptron, I understood how error gradients propagate from the output back to the input, and how weights ($w$) and biases ($b$) are adjusted using the learning rate ($n$).

3. Activation Functions & Derivatives: I manually coded the mathematical formulas for Sigmoid, Tanh, and ReLU, along with their specific derivatives needed for the gradient descent step. I learned why ReLU is crucial for preventing vanishing gradients in deeper networks.

4. NumPy Vectorization: Handling matrix multiplication (np.dot) and shape alignment was challenging. I learned how to efficiently process inputs and weights as vectors rather than looping through every single value.

5. Model Serialization (The .dit Format): I learned how to persist state. By implementing the save() and load() methods, I figured out how to extract weights/configs from objects, store them in a binary format (.npz), and reconstruct the exact object architecture from that file later.

💻 Usage Example

import aino

model = aino.NeuralNetwork([2, 3, 1], activation_type='sigmoid')
model.fit(X_train, y_train, epochs=100, n=0.1)
model.predict(X_test)
model.save('aino.dit')

Built with ❤️ use Python and Numpy.