Principle:Snorkel team Snorkel Slice Aware Classification
| Knowledge Sources | |
|---|---|
| Domains | Data_Slicing, Multi_Task_Learning, Robustness |
| Last Updated | 2026-02-14 20:00 GMT |
Overview
A multi-task learning approach that augments a base classifier with per-slice indicator and predictor heads to improve performance on critical data subsets.
Description
Slice-Aware Classification addresses the problem of maintaining robust model performance across critical data subsets. Instead of training a single classifier that may underperform on important slices, this approach creates additional task heads for each slice:
- Indicator heads: Binary classifiers that predict whether a data point belongs to each slice
- Predictor heads: Per-slice prediction heads that specialize on slice-specific patterns
- Master head: A combining head that uses attention-based reweighting (SliceCombinerModule) to aggregate slice-specific representations
This architecture allows the model to learn slice-specific representations while maintaining overall performance through shared feature extraction.
Usage
Use this principle when you have identified critical data slices and want a single model that performs well both overall and on each important slice. It is most effective when slices have distinct patterns that benefit from specialized prediction heads.
Theoretical Basis
The slice-aware architecture creates a multi-task learning problem where the model simultaneously optimizes:
The final prediction combines slice representations through learned attention weights:
Failed to parse (syntax error): {\displaystyle \hat{y} = \text{master\_head}(\sum_{j} \alpha_j \cdot h_j) }
where are attention weights and are slice-specific representations.