Learning invariances via correlation and marginal alignments for out-of-distribution generalization

Key Points

• Out-of-distribution generalization improves with invariant learning mechanisms, showing significant advantages.
• Models utilizing correlation and marginal alignments yield better performance, achieving up to 30% increased accuracy.
• Observational analysis using synthetic datasets reveals that structured alignments effectively capture variability.
• These findings highlight the importance of invariance strategies for creating more robust models against data shifts.