Which phrase describes a solution that seems good but isn't truly optimal, requiring overcoming a short dip?

Which phrase describes a solution that seems good but isn't truly optimal, requiring overcoming a short dip?

• A. Local Optima
• B. Plateaus
• C. Vanishing Gradients
• D. Exploding Gradients

Answer: (A)

Local optima occur when a solution is the best within its immediate neighborhood but not the best overall. In many problems, the landscape has many hills and valleys, so you can find a point that looks good and seems to be performing well locally. However, small moves around that point don’t improve the objective, making it feel like you’re done when you’re not. To reach the truly optimal solution, you often need to move through a region where the objective dips temporarily, then climb to a better value in another area. That’s exactly what overcoming a short dip describes. Plateaus are flat regions with little or no gradient, which slows progress rather than indicating suboptimality in the sense of being stuck away from the global optimum. Vanishing gradients and exploding gradients refer to problems where gradient magnitudes shrink to near zero or explode to large values during training, respectively, hindering learning rather than describing a suboptimal local solution that requires dipping to improve.

Local optima occur when a solution is the best within its immediate neighborhood but not the best overall. In many problems, the landscape has many hills and valleys, so you can find a point that looks good and seems to be performing well locally. However, small moves around that point don’t improve the objective, making it feel like you’re done when you’re not. To reach the truly optimal solution, you often need to move through a region where the objective dips temporarily, then climb to a better value in another area. That’s exactly what overcoming a short dip describes.

Plateaus are flat regions with little or no gradient, which slows progress rather than indicating suboptimality in the sense of being stuck away from the global optimum. Vanishing gradients and exploding gradients refer to problems where gradient magnitudes shrink to near zero or explode to large values during training, respectively, hindering learning rather than describing a suboptimal local solution that requires dipping to improve.