Choose a loss that matches the task and the metric you care about, and understand what each penalises. Use when a model optimises well and performs badly on the thing that matters.
