AI Logbook

Why do we wrap our math inside a class LinearRegressor instead of just running a loose script? Statefulness and Data Leakage.

Why do neural networks demand scaled data? It comes down to the geometry of the Loss Landscape.

When we trained on a single house, the gradients only pointed toward the perfect answer for that specific house.

Why do we calculate two separate gradients? Because $w$ and $b$ do entirely different things geometrically.

This concept of multiplying derivatives backward through a chain of equations is called Backpropagation.

This concept is the core intelligence of every neural network in existence, from the simplest regressor to Large Language Models. It is called Gradient Descent.

Why do we go through all the trouble of building matrices? Parallelization.

Why is linear dependence a problem for Artificial Intelligence? In data science, this concept is called Multicollinearity, and it destroys models.

Why is transposition so critical in machine learning? It is the ultimate "adapter cable" for matrix multiplication.

Why do we need to multiply two matrices together? This is the exact mathematical operation that unlocks Deep Learning.

In The Capstone: Object-Oriented ML Architecture, to make predictions for 3 houses, we had to run our predict method 3 separate times. In Python, this requires a for loop, which is notoriously slow for large datasets.

Why do we need these operations? In machine learning, we rarely use raw data exactly as it comes to us. We need to preprocess it so our neural networks can digest it effectively.

In Week 1, we built an engine that could look at a single house (a vector) and predict a price. But in the real world, we don't train AI on one house at a time; we train it on thousands or millions of houses simultaneously.