Stochastic Gradient Descent vs Subgradient Methods

Developers should learn SGD when working with large-scale machine learning problems, such as training deep neural networks on massive datasets, where computing the full gradient over all data points is computationally prohibitive meets developers should learn subgradient methods when working with optimization problems involving non-differentiable convex functions, such as in training support vector machines or solving large-scale linear programs. Here's our take.

Stochastic Gradient Descent

Nice Pick

Developers should learn SGD when working with large-scale machine learning problems, such as training deep neural networks on massive datasets, where computing the full gradient over all data points is computationally prohibitive

Pros

• It is particularly useful in online learning scenarios where data arrives in streams, and models need to be updated incrementally
• Related to: gradient-descent, optimization-algorithms

Cons

• Specific tradeoffs depend on your use case

Subgradient Methods

Developers should learn subgradient methods when working with optimization problems involving non-differentiable convex functions, such as in training support vector machines or solving large-scale linear programs

Pros

• They are particularly useful in machine learning for handling L1 regularization (e
• Related to: convex-optimization, gradient-descent

Cons

• Specific tradeoffs depend on your use case

The Verdict

These tools serve different purposes. Stochastic Gradient Descent is a methodology while Subgradient Methods is a concept. We picked Stochastic Gradient Descent based on overall popularity, but your choice depends on what you're building.

The Bottom Line

Stochastic Gradient Descent wins

Based on overall popularity. Stochastic Gradient Descent is more widely used, but Subgradient Methods excels in its own space.