Skip to content

Multi-Task Learning

Definition

Multi-task learning (MTL) trains a single model on multiple related tasks simultaneously — sharing representations across tasks for better generalization and efficiency.

MTL in eKYC

graph TD
    A[Shared Backbone<br/>ResNet-18 / ViT-S] --> B[Liveness Head<br/>Real vs Spoof]
    A --> C[Depth Head<br/>Face depth map]
    A --> D[Domain Head<br/>Which dataset/domain]
    A --> E[Quality Head<br/>Face quality score]

    style A fill:#4051B5,color:#fff
Task Combination Benefit
Liveness + depth estimation Depth provides geometric reasoning about 3D vs 2D
Liveness + domain classification Adversarial domain head forces domain-invariant features
Recognition + quality estimation Quality awareness improves matching on low-quality inputs
OCR detection + recognition Shared features improve both stages
Document classification + field extraction Document type informs field locations

Key Takeaways

Summary

  • MTL improves liveness accuracy by 2-5% over single-task (binary) training
  • Liveness + depth + domain is the standard multi-task combination for face anti-spoofing
  • Shared backbone reduces total compute — one model serves multiple purposes
  • Task weighting (relative loss importance) requires careful tuning