Multi-Task Learning
Training one shared network on several related tasks at once, so what one task teaches the shared layers can help the others, for one inference cost instead of several.
Hard parameter sharing runs one shared trunk of layers across every task, with a separate small head per task, so each task's gradient shapes the shared layers while only its own head stays private. This helps when tasks are related enough that a representation good for one also helps the others.
Negative transfer happens when tasks pull the shared trunk in conflicting directions, visible directly as negative cosine similarity between their gradients on the shared layers.