Syllabus Map

• Study map: Syllabus Study Map

Core Idea

• Gradient is the vector of partial derivatives for scalar output.
• Jacobian generalizes derivatives to vector outputs.

Key Formulas

Practical Notes

Negative gradient gives steepest local decrease.

• This is the basis of gradient descent updates.

Jacobians are useful for sensitivity and transformations.

• Especially relevant in sequence and manifold models.

Why This Matters for ML

• Gradient vectors determine direction and strength of optimization updates.
• Jacobians describe how representations change under transformations, useful in sensitivity checks.
• Many modern objectives (contrastive, sequence, diffusion components) rely on multivariate gradients.
• Understanding gradient geometry helps with clipping, normalization, and optimization stability.