GO4Align: Group Optimization for Multi-Task Alignment

Summary

The paper designs a multi-task optimization method, namely GO4Align, to address task imbalance by aligning optimization processes across tasks. It proposes an adaptive group risk minimization strategy, formulated as a bilevel optimization problem where the lower-level optimization is a task grouping optimization and upper-level optimization is a weighted optimization over task group losses. Then, this method optimizes the problem by alternating two steps: (1) Dynamical Group Assignment where tasks are clustered using a dynamic group assignment process, implemented via K-means clustering, to capture beneficial task interactions. (2) Risk-guided Group Indicators: indicators are designed to balance task risks and align learning progress by combining scale-balance and smooth-alignment operations. GO4Align is evaluated on benchmarks, including NYUv2, CityScapes, QM9, and CelebA. The results show that it outperforms existing gradient-oriented (MGDA, PCGRAD, CA-GRAD, IMTL-G, GRADDROP, and NASHMTL) and loss-oriented methods (Linear scalarization, Scale-invariant, Dynamic Weight Average, Uncertainty Weighting, Random Loss Weighting, and FAMO.), achieving lower performance drops and better computational efficiency. Lastly, the study explores the contributions of each component of GO4Align, the influence of group assignments, and the role of group weights. It shows that the proposed AGRM principle can integrate with existing MTO methods, further improving their performance.

Strengths

- This paper proposes an adaptive group risk minimization strategy to address task imbalance, formulated as a bilevel optimization problem where the lower-level optimization is a task grouping optimization and upper-level optimization is a weighted optimization over task group losses. - GO4Align is evaluated on benchmarks, including NYUv2, CityScapes, QM9, and CelebA. The results show that it outperforms 12 existing gradient-oriented and loss-oriented methods, achieving lower performance drops and better computational efficiency.

Weaknesses

- There are many existing clustering algorithms that can be potentially used for task grouping, such as spectral clustering and SDP-based clustering [1]. What is the rationale for choosing K-means clustering in the method? It would be better to discuss and ablate the clustering algorithms, as it is an important component of the proposed method. - Another important component is the choice of the group indicators for clustering tasks. This work uses indicators based on task loss trajectories. How about using gradients and model features as group indicators? Would it be worse than the proposed indicators in Section 3.3? - It would be better to explain why the proposed method uses the comparable time as linear scalarization. How does the proposed method scale the number of tasks? It would be better to provide a comparison of the additional computation across each method? [1] Relax, no need to round: integrality of clustering formulations. https://arxiv.org/abs/1408.4045

Questions

- How is the convergence difference evaluated in Figure 2? How is the $\Delta m$ metric defined?

Rating

7

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

This work has discussed its limitations.

Summary

This paper proposes a multi-task optimization method designed to address task imbalance by aligning optimization processes across different tasks. To accomplish this, the authors developed an adaptive group risk minimization strategy, which includes two key techniques: (i) dynamic group assignment, clustering similar tasks based on their interactions, and (ii) risk-guided group indicators, leveraging consistent task correlations and risk information from previous iterations. Extensive experimental results across various benchmarks show that the proposed method outperforms others while requiring even lower computational costs.

Strengths

1. The paper is well-organized and easy to follow. 2. The proposed method outperforms several newly proposed MTO algorithms. 3. Extensive experimental results across various benchmarks show that the proposed method outperforms others while requiring even lower computational costs.

Weaknesses

1. To help readers better understand the application scenarios of the proposed method, the author can provide examples illustrating the phenomenon of task imbalance in MTO tasks. 2. In Eq. (4), what is the optimization method for $w$ and $G$? Is it performing standard k-means on $\gamma$? 3. In Line 172, why is it possible to invert the $K\times M$ matrix $G$? 4. In the introduction of the dataset, there is a lack of explanation regarding the imbalance phenomenon in the relevant learning tasks.

Questions

See "Weaknesses".

Rating

6

Confidence

3

Soundness

2

Presentation

3

Contribution

3

Limitations

No.

Summary

The paper proposes GO4Aligh, a multi-task optimization approach, which targets the task imbalance issue. Specifically, the authors devise two objectives in multi-task optimization: 1) the first is the dynamical group assignment, which can attribute similar tasks to the same cluster and distinguish different ones. 2) The second is the risk-guided group indicators, which include scale-balance and smooth-alignment. These two objectives are incorporated into the AGRM principle for multi-task optimization.

Strengths

From an overall view, this paper is quite well-structured, with clear writing and an organized layout. The primary motivation, which is to optimize task imbalance, is clearly articulated and valuable. The methodology is understandable and easy to reproduce. Also, the experiments show that GO4Align can reduce 6.08% metrics with no further training time.

Weaknesses

While I do not have major concerns about this paper, it’s worth noting that the performance improvement of GO4Align over FAMO is marginal, which may limit its broader impact. The concept of grouping-based task interactions is innovative, but not necessarily groundbreaking. Besides, I have some questions. Please refer to the next part.

Questions

- Given the existence of Eq.4, I’m unsure whether Eq.5 or Eq.6 would have a significant impact. How would the performance change if only Eq.4 is included? - The paper asserts that GO4Align has even lower computational costs, yet it is 1.02x that of STL. Does this mean that GO4Align is more efficient compared to the baselines (not the STL)?

Rating

7

Confidence

3

Soundness

3

Presentation

4

Contribution

2

Limitations

Yes.

Summary

This paper presents an approach for multi-task learning aiming to reduce interference among tasks by adaptive loss weighting. Instead of task-specific weights as in existing works, the authors proposed to group similar tasks and all tasks in the same group share the same weight. The paper also includes a new derivation of task level weights (referring to group indicator), which contains two components: scale-balance and smooth-alignment. Evaluation was done using four benchmark datasets, three in computer vision and one in chemistry (predicting property of molecules)

Strengths

• The manuscript is very well written and easy to follow • The proposed idea is neat. The proposed task grouping is flexible, which can easily work with other task-weight deriving methods. • Ablation study was done to evaluate the contribution of ingredients in the proposed method.

Weaknesses

• I see that the technical contributions of this work are twofold. One is the derivation of the group indicator (essentially task specific weights) and the other one is to use the same weight to weigh all tasks in a group instead of just using task specific weights. I do not see clear motivation of both. More discussion is needed why both are expected to be better than existing approaches. • Figure 5 needs more description to make it easier to read. It took me quite some effort and time to figure the x-axis in the plots are epochs and the intensity of the color indicates the weight value (if I am correct). • There are results from only one dataset which contains just three tasks. Results from more datasets especially those with larger number of tasks are desired, given the proposed method aims at task grouping. • The reported results are not strong. (1) Based on the results from CityScapes (table 2), which has only two tasks (meaning there is not much grouping involved and difference in performance should be due to that in weight calculation), GO4Align is not the best performing one, implying the proposed weight calculation could be inferior to existing one. (2) All numbers in Table 2 are positive, implying models perform worse than those in single-task learning, which defeats the major advantage of MTL, enhancing the model generalizability and performance. • I do not see an ablation study that compares with grouping to without grouping at all. • More details in experiment setup is needed to enhance reproducibility, especially, for each of the experiment how partition of the dataset was done for training, validation, and testing.

Questions

Refer to weakness

Rating

5

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

NA