The regression framing threw me a bit since I was expecting classification.
Start by acknowledging the challenge of differing scales and signs, then propose a multi-task learning architecture with task-specific normalization and loss weighting. Discuss techniques like uncertainty weighting or gradient normalization to balance tasks, and mention evaluation strategies to ensure both tasks improve.
Pro tip: Emphasize that the goal is not just to balance losses but to ensure both tasks benefit from shared representations; consider using techniques like GradNorm or dynamic weight adjustment based on validation performance.
Recognize that the tasks have different scales (one negative, one positive) and potentially different units, which can cause one task to dominate training. Also note that negative values can complicate normalization and loss functions.
Apply task-specific normalization: for the positive task, use standard scaling; for the negative task, consider shifting to positive or using robust scaling. Alternatively, standardize each task's targets to zero mean and unit variance.
Use a shared base network with task-specific heads. This allows shared representations while accommodating different output distributions. Consider separate output layers with appropriate activation functions (e.g., linear for regression).
Combine losses with weights. Use static weights tuned via cross-validation, or dynamic methods like uncertainty weighting (Kendall et al.) or GradNorm to automatically balance tasks during training.
Monitor each task's performance on validation sets. Use metrics like per-task loss and correlation of predictions. Adjust weighting or architecture if one task lags. Consider multi-task learning benefits like improved generalization.
AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.