Task-Aware Dynamic Model Optimization for Multi-Task Learning

📄 IEEE Access 2023 — DOI: 10.1109/ACCESS.2023.3339793

This repository introduces the core ideas and implementation of
Task-Aware Dynamic Model Optimization (DMO), a memory-efficient multi-task learning (MTL) framework.

---

🔍 Overview

Multi-task learning (MTL) often suffers from task interference and inefficient resource usage.
DMO (Dynamic Model Optimization) addresses these issues by:

• Grouping tasks based on inter-task similarity (weight & loss correlations)
• Dynamically allocating parameters based on task difficulty
• Producing lightweight subnetworks, reducing parameters by over 80% on average

_{Figure: Overview of Task-Aware Dynamic Model Optimization (DMO)}

---

🎯 Key Components

• Task Similarity Measurement
  • Pearson correlation on weights (S_W)
  • Pearson correlation on losses (S_L)
• Task Grouping
  • Graph-based maximal clique detection (Bron–Kerbosch)
• Dynamic Parameter Allocation
  • Pruning by task difficulty
  • Merging into group-specific subnetworks

---

🛠️ Code Structure

Task-Aware-DMO/
├── models/           # Backbone network, task-specific heads, pruning utilities
├── train.py          # Training logic
├── evaluate.py       # Evaluation and metrics
├── config/           # Experimental settings
├── results/          # Graphs and tables from the paper
├── assets/           # Images and visual materials for README
└── paper/            # Original PDF paper for reference

---

🧪 Datasets Used

• CIFAR-10, STL-10, MNIST, USPS
• CIFAR-100 (20-task version)
• Visual Decathlon Challenge (10 diverse tasks)

---

📊 Result Summary

Dataset Group	Param Reduction	Accuracy (Avg)
CIFAR-10 + MNIST	~84%	⭤ Comparable
CIFAR-100 (20 Tasks)	~85%	⬆️ Improved
Decathlon Challenge	~88%	⭤ Slight Drop

🔹 Up to 97% reduction on USPS with minimal performance loss

---

🧪 Visual Decathlon Challenge (Summary Table)

Method	Avg Accuracy (%)	# Parameters (M)	Param Ratio
Single-task	74.69	101.68	1.00
Soft Param Sharing [7]	65.24	101.68	1.00
Cross-stitch [8]	76.33	203.36	2.00
NDDR [10]	59.76	252.86	2.49
TAPS [6]	59.11	130.50	1.28
DMO (Ours)	74.40	11.48	0.11

---

📑 Visual Decathlon Table (From Paper)

_{Figure: Visual Decathlon Challenge results table (from IEEE Access paper)}

---

🧑‍💻 Author

Su-jin Choi (최수진)
Master of AI, Chung-Ang University
📧 [email protected]
📄 IEEE Access Paper