Скачать или смотреть Coding a Neural Net from Scratch (The Math You Need)

Stop treating Neural Networks like a black box. In this video, we build a fully connected Neural Network completely from scratch using only Python and NumPy.

We don't just type out code, we derive the math first. I walk you through the chain rule, how to handle matrix shapes during backpropagation, and why we use the outer product for weight updates. By the end, we train our custom network on a complex non-linear dataset to prove the math actually works.

📂 Get the Code (GitHub): https://github.com/amoughnieh/numpy-n...
(Note: The repo also includes a bonus section on Binary Cross Entropy and classifying a Spiral Dataset!)

Timecodes:
00:00 - Intro
00:40 - The Math: NN as a Function Approximator
01:49 - Gradient Descent & The Loss Landscape
03:04 - Deriving the Chain Rule
08:04 - Forward Pass: Linear Algebra & Dimensions
14:11 - Why We Need Non-Linear Activations (Sigmoid)
16:20 - Designing the OOP Structure
19:55 - Coding: The Linear Layer
22:40 - Coding: Sigmoid & MSE Loss
27:54 - Backpropagation Implementation
34:00 - Handling Matrix Shapes (Transposing)
43:00 - The Outer Product: Solving Weight Updates
49:20 - Training Loop & First Results
52:48 - Final Test: Fitting a Complex Sinusoidal Wave

Key Concepts Covered:
Forward Pass: Linear combinations and activation functions.
Backpropagation: Calculating partial derivatives using the Chain Rule.
Matrix Calculus: Understanding why we transpose matrices during the backward pass.
Optimization: Implementing Gradient Descent without external libraries.

#Python #MachineLearning #DeepLearning #NumPy #NeuralNetworks #DataScience