Small Batch (32): Noisy updates, slower but better generalization
Medium Batch (128): Balanced performance
Large Batch (512): Stable updates, faster but may overfit