Multi-task and multi-objective

How this question breaks down

1. 1Task headsCore17 minOne network can learn several predictions at once, with shared embeddings and layers at the bottom, one small tower per task on top, and a loss that adds the tasks up. Sparse tasks borrow what the busy ones learn, and serving pays for the bottom once. Mixture-of-experts variants let each task choose what to share.
2. 2BlendingCore17 minA ranker with five heads still has to show one list, so the heads' predictions are combined into one score, as a weighted sum, a weighted product, an expected value or a hard filter. The weights are a product decision, not a training detail. They only mean something if the heads are calibrated, and teams set them by sweeping them in live experiments along the trade-off frontier.
3. 3ConflictsCore17 minSharing helps until one task's gradient drowns out or points against another's, and a task ends up worse than it would be alone. Diagnose it by comparing each head with its single-task version and by looking at loss scales and gradient directions. Fix it in order: rescale and tune the loss weights, then change the architecture, then split the task out.

Work through the parts in order and you’ve covered the full design. Each part is one topic.