GO4Align: Group Optimization for Multi-Task Alignment

TL;DR

Comprehensive experimental results on diverse benchmarks demonstrate the proposed Go4Align method's performance superiority with even lower computational costs.

Abstract

This paper proposes \textit{GO4Align}, a multi-task optimization approach that tackles task imbalance by explicitly aligning the optimization across tasks. To achieve this, we design an adaptive group risk minimization strategy, comprising two techniques in implementation: (i) dynamical group assignment, which clusters similar tasks based on task interactions; (ii) risk-guided group indicators, which exploit consistent task correlations with risk information from previous iterations. Comprehensive experimental results on diverse benchmarks demonstrate our method's performance superiority with even lower computational costs.

Figures (8)

• Figure 1: Performance and computational efficiency evaluation for MTO methods evaluated onNYUv2. Each method's training time is relative to a baseline method, which minimizes the sum of task-specific empirical risks. Left-bottom marks comprehensive optimal results.
• Figure 2: Multi-task alignment and effects on performance. We visualize relative task performance curves (lower is better) over training epochs. Better overall performance usually occurs with lower convergence differences. Our method effectively reduces the convergence difference and achieves a better overall performance.
• Figure 3: GO4Align using adaptive group risk minimization in the bi-level optimization framework. In the lower-level optimization, the model assigns tasks to groups with different group weights, encouraging task interactions and aligning learning progress. Such group information is nested into the upper-level optimization for updating the multi-task model's parameters.
• Figure 4: Efficiency comparisons on training time. Each method’s training time is relative to a simple baseline method with Eq. (\ref{['eq: scale-ERM']}), which minimizes the sum of task-specific empirical risks.
• Figure 5: Comparative analysis of the influence of the group assignment matrix and group weights onNYUv2. The x-axis in the subplots denotes the epoch, and the intensity of the color indicates the weight value.(a-d) have fixed group weights $\bm{\omega}=[\omega^1, \omega^2]$ but various group assignment matrices $\mathcal{G}$. (f-i) have various group weights $\bm{\omega}$ but a fixed group assignment matrix $\mathcal{G}$. (e) is our method that dynamically exploits a group assignment matrix and group weights for each iteration. The right side of each method shows relative performance drops on each task and their average one.