I stared at this longer than I'd like to admit.
Start by clarifying the problem: one task has negative labels with a wide range, the other positive labels with a different scale. Then propose a multi-task learning architecture with separate output heads and a shared backbone, using normalization or loss weighting to handle scale differences, and discuss trade-offs like task interference and evaluation metrics.
Pro tip: Emphasize that you would monitor per-task performance and use techniques like uncertainty weighting or GradNorm to dynamically balance losses, showing awareness of practical multi-task training challenges beyond simple normalization.
Clarify the label distributions, ranges, and whether the tasks are related. Consider if negative labels are truly a separate task or a transformation of the same underlying quantity.
Propose a shared neural network backbone with two separate output heads: one for the negative-range task and one for the positive-range task. This allows shared representation learning while accommodating different output scales.
Normalize each task's targets (e.g., standardize or min-max scale) before training, and use appropriate loss functions (e.g., MSE on normalized targets). Alternatively, use loss weighting to balance contributions.
Employ techniques like uncertainty weighting, GradNorm, or dynamic weight averaging to prevent one task from dominating. Mention that simple loss weighting may require tuning.
Monitor per-task metrics on a validation set, check for negative transfer, and consider techniques like task sampling or separate batch normalization if needed. Discuss potential need for task-specific learning rates.
AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.