Curriculum Learning in Multi-Task Deep Learning

Curriculum Learning in Multi-Task Deep Learning

Introduction

Curriculum learning—training models on progressively harder tasks—has proven effective for single-task learning. When extended to multi-task learning scenarios, curriculum strategies become even more powerful but also more complex to design effectively.

Multi-Task Learning Fundamentals

Multi-task learning trains a single model on multiple related tasks simultaneously. Benefits include:

• Shared representations reduce parameters
• Transfer learning between tasks
• Better generalization through regularization
• Computational efficiency

However, naive MTL often suffers from:

• Task interference (one task hurts another’s performance)
• Unbalanced convergence rates
• Gradient conflicts during backpropagation

Curriculum Learning Applications

Task Sequencing

Rather than training all tasks equally from the start:

1. Easy-to-Hard: Start with simpler, better-labeled tasks
2. Related-to-Specialized: Begin with general tasks, move to specific ones
3. Prerequisite-Based: Order tasks by their dependencies

Instance-Level Curriculum

For each task, apply difficulty curriculum:

• Start with clean, representative examples
• Gradually introduce harder/noisier samples
• Adapt difficulty based on per-task loss

Joint Curriculum Strategies

• Interleave task training based on convergence status
• Dynamically weight tasks based on learning progress
• Detect and prevent task interference

Implementation Strategies

Metric-Based Scheduling

for epoch in epochs:
  for task in tasks:
    if task_loss[task] > threshold:
      increase_difficulty_curriculum(task)
      weight[task] = higher_weight
    else:
      weight[task] = lower_weight

Pacing Functions

• Self-paced learning: Model selects which samples to train on
• Teacher-paced learning: External curriculum guides the process
• Mixed-paced learning: Hybrid of both approaches

Benefits and Trade-offs

Advantages

• Faster convergence compared to uniform training
• Better final performance on harder tasks
• Reduced task interference
• More stable gradient flow

Challenges

• Curriculum design is task-specific
• May require manual tuning
• Computational overhead of tracking per-task progress
• Risk of getting stuck in local minima

Real-World Examples

Computer Vision: Object Detection + Classification

• Start with classification (simpler, fully-labeled data)
• Add detection (uses classification as foundation)
• Gradually introduce rare classes

NLP: Named Entity Recognition + POS Tagging

• Begin with POS tagging (simpler linguistic task)
• Then NLP with entity relationships
• Difficulty increases with out-of-domain data

Robotics: Imitation Learning + Reinforcement Learning

• Curriculum from supervised learning to RL
• Gradual policy autonomy increases

Research Directions

• Automatic curriculum discovery through meta-learning
• Theoretical understanding of task interference
• Optimal ordering strategies for many tasks
• Balancing exploitation vs. exploration in curriculum design

Practical Tips

1. Start Simple: Begin with single-task curriculum learning first
2. Monitor Gradients: Watch for conflicting gradient signals
3. Adaptive Weights: Use loss-based or progress-based weighting
4. Validation Matters: Test on held-out data from all tasks regularly
5. Document Trade-offs: Different curricula may excel at different tasks

References

• Curriculum Learning papers and implementations
• Multi-task learning surveys
• Gradient conflict analysis literature
• Real-world case studies in curriculum-based MTL