ML Prerequisites Part 3: Feature Scaling Explained | Normalization vs Standardization | Tamil ML
📌Github: https://github.com/Ganesh2609/TensorT...
This video is ML Prerequisites Part 3, continuing the structured learning path after bias, variance, and overfitting. In this part, we focus on one of the most practically important topics in Machine Learning: Feature Scaling.
We begin by revisiting the problem setup—what Machine Learning models actually solve and why numerical scale matters. Since ML models operate purely on numbers and mathematical operations, feature ranges play a critical role in how models learn. To build intuition, we briefly connect this idea to how tokenizers and transformer models convert information into numerical representations.
Using a simple and relatable example, such as age ranging from 0–100 versus salary ranging up to lakhs, we show how features on different scales affect optimization. You will clearly see how larger-scale features dominate gradient descent updates, leading to slower convergence, unstable training, and poor model performance.
From this motivation, we formally introduce feature scaling in Machine Learning and then zoom out to the bigger picture, where two major techniques emerge: normalization and standardization.
We explain what normalization is and how Min–Max Scaling works, what standardization is and how the Z-score or Standard Scaler works, and what outliers are and how each method behaves in the presence of outliers.
To strengthen intuition, we provide clear numerical and visual explanations before moving into implementation.
On the coding side, we implement a Min–Max Scaler from scratch to understand the math, use sklearn’s MinMaxScaler while explaining fit, transform, and fit_transform, explain why scalers must be fit only on training data and never on test data, show why test data can have values outside the 0–1 range after scaling, and implement and explain the StandardScaler.
Finally, we use histograms to visually compare raw data, min–max scaled data, and standardized data, reinforcing when and why each technique should be used.
This video is a mandatory prerequisite before implementing models like Linear Regression, Logistic Regression, SVMs, and Neural Networks.
📌Chapters
00:00 - Introduction
00:15 - The problem, why we use feature scaling
03:57 - What is feature scaling
04:35 - The bigger picture: Standardization and Normalization
10:46 - Code implementation: custom Min-Max Scaler
14:35 - Min-Max Scaler using sklearn
19:35 - Standard Scaler
21:44 - Graphical explanation with histogram
📌Keywords
ml prerequisites part 4 tamil, feature scaling tamil, normalization vs standardization tamil, min max scaler tamil, standard scaler tamil, gradient descent scaling tamil
📌Tags
MachineLearning, MLPrerequisites, FeatureScaling, Normalization, Standardization, TamilML, DataScienceTamil, AITamil
📌GitHub Repository (All Codes):
https://github.com/Ganesh2609/TensorT...