What do you understand by 'feature scaling', and why is it important?
Instruction: Define feature scaling and its importance in the context of training deep learning models.
Updated
Example Answer
Feature scaling puts numeric inputs on useful, comparable scales. For standardization, I subtract the training-set mean and divide by the training-set standard deviation. The training feature then has mean zero and unit variance when its variance is nonzero. This does not make a skewed or bimodal distribution Gaussian.
Scaling matters for distance-based models, regularized linear models, and many neural-network optimization problems. A feature measured in thousands should not dominate just because another is measured in fractions. I fit the scaler on training data only, then reuse those statistics on validation, test, and production inputs. Refitting on the test set would leak information and change the input representation.
I would inspect outliers before choosing a scaler. Standardization is sensitive to them; a robust scaler may be more appropriate. Constant features also need explicit handling rather than division by zero.
Worked Example
For training values [1, 1, 1, 9], the mean is 3 and the population standard deviation is √12. Their standardized values are approximately [-0.577, -0.577, -0.577, 1.732]: centered and scaled, but still strongly asymmetric.