acceptodds
Withdrawn from ICLR 2027

Can Control Learning with Bandit Optimization Accelerate Convergence in Multimodal Fusion of Medical Data?

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Multimodal fusion has become a central paradigm in medical AI, integrating heterogeneous sources such as imaging and multi-omics for diverse downstream tasks. However, its training-time efficiency remains underexplored. As training progresses, modalities and tasks contribute unevenly to learning, yet prior methods often optimize the shared multi-task loss using fixed or heuristic open-loop weighting. This mismatch limits adaptation to evolving learning dynamics and can delay convergence to strong validation performance. To address this challenge, we propose Bandit with Control-Learning Aware optimizatioN (BCLAN), a novel parameter-free hybrid closed-loop controller that augments standard optimizers without architectural changes or additional trainable backbone parameters. Driven by validation feedback, BCLAN couples adversarial bandit exploration over a compact set of loss-weight templates with proportional–integral–derivative (PID)-based refinement in logit space, enabling it to search for improved task– modality allocations while stabilizing their adaptation under noisy, non-stationary feedback. We provide theoretical insights into this coupling, establishing bounded loss-weight refinement and identifying conditions for faster convergence. Across twelve heterogeneous-modality benchmarks in the medical domain, BCLAN improves downstream performance by up to ≈ 2.7% and reduces training time by up to ≈ 50.4% compared with the corresponding baselines, achieving a favorable training-time efficiency–performance trade-off.

Primary area:
optimization
Keywords:
Optimization, Multimodal fusion, Medical data analysis

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_ppd9OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper argues that fixed weighting of task- and modality-specific losses delays convergence in multimodal medical models. It proposes BCLAN, a plug-in controller that reweights losses at each validation interval: an EXP3-IX bandit samples a weight template from a small fixed pool, and a PID controller adds a bounded logit-space bias driven by validation feedback. Theorem 1 bounds the weight deviation and expected time to reach a validation-loss threshold; Corollary 1 gives a condition for faster convergence than static weighting. On 12 datasets with seven host models and four online adaptive baselines, the paper reports up to 2.7-point gains and 30–50% lower per-epoch training time for the two-head, supported by ablations and sweeps.

Critical strengths
  1. Relevant problem and lightweight plug-in design. The paper moves the efficiency question in multimodal medical fusion from architecture to loss allocation during training, and BCLAN wraps existing pipelines without architectural changes or additional trainable parameters, adding about 1.4% controller overhead per epoch.
Critical weaknesses
  1. The central claim of faster convergence is untested, and the evaluated controller is not the analysed one. The paper defines its objective as the expected time to reach a validation-loss threshold (Eqs. 4–5) but never reports this time, or the threshold, for any method; it reports per-epoch training time (#TTE) instead. Loss reweighting leaves per-epoch computation unchanged and the controller adds overhead (Table 16), so the 30–50% #TTE reductions lack a mechanism.

  2. The reported gains cannot be attributed to closed-loop control. All competing methods are sampling, learning-rate, or calibration schemes rather than loss-weighting methods; adaptive weighting baselines (e.g., uncertainty weighting, GradNorm), a validation-selected static weight, and random template switching are missing.

Questions for authors
  1. How is #TTE measured, and how does loss reweighting reduce per-epoch time by 30–50% when architecture, FLOPs, data, and batch size stay unchanged? Did training use early stopping? Please report the paper’s own objective—epochs and wall-clock time to a pre-specified threshold J*—for every method and seed, and state J* for each dataset.

  2. Which controller produced Tables 1–4: the loss-based version of §2.3, the accuracy-based version of App. E (A = 5)? Table 15 reproduces the main result at A = 8. Please reconcile these settings and report results for the loss-based controller that Theorem 1 covers.

  3. Table 2 repeats identical standard deviations across different methods (e.g., ±7.38/±2.38/±3.43 for four DRIFA-Net rows; MAIL+ADAMBS and MAIL+Incremental PID share all three values). How were these computed? Please provide per-seed or per-split results for Tables 1–3, standard deviations for Table 1, and paired differences with confidence intervals or paired tests.

  4. How was the “oracle fixed” template in Table 15 selected (validation or test)? Please add a validation-selected static weight, uniform random template switching without bandit updates, adaptive weighting methods (uncertainty weighting, GradNorm, DWA, FAMO), and multimodal balancing methods (e.g., OGM-GE, AGM, MMPareto). The current competitors address sampling, learning-rate scheduling, or calibration, not loss weighting.

  5. Please plot λ, b, and arm probabilities over training. Because b accumulates the full PID output (Eqs. 15, 47), the gains act on b as Kd·ε+Kp·Σε+Ki·ΣΣε, and B=2 permits a ±4 shift in the log-weight ratio, whereas the templates span only ±1.39. Does b saturate? What is d in §2.3? Does the bandit favour better templates when its reward mainly tracks training progress?

  6. HAM10000 accuracy of 99.4–99.6% and MIMIC-III mortality AUC up to 95.9% exceed values commonly reported for these benchmarks. Are splits lesion- and patient-level (HAM10000 contains several images per lesion)? The released code subsamples the test set (810 of 1001), splits SIPaKMeD 64/16/20 rather than the stated 80/10/10, and keeps dropout active at test time; please clarify.

  7. Fig. 5 sweeps a “reward scale” that appears in no equation and cannot apply to Incremental PID or PID-AGD; Table 16 uses an undefined T̂*_R. Please define or remove these, report sensitivity to Kp, Ki, Kd, B, λ_b, the target, and the template values, and justify “parameter-free”.

  8. Evaluate BCLAN on genuinely paired multimodal data with more than two loss components, and specify how to construct templates when T·m > 2.

Reviewer_SmUZOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Summary

This paper proposes BCLAN, which combines bandit optimization and PID control for adaptive loss weighting in medical multimodal learning. The method is evaluated on multiple medical datasets and shows improvements in both performance and training efficiency. Overall, I think this is an interesting work with meaningful motivation.

Critical strengths

Strengths

  1. I think improving training efficiency is an important problem, especially for medical multimodal learning, where training can be computationally expensive.
  2. The combination of bandit optimization and PID control is interesting. It's also interesting to find that the method can be applied to existing models without changing the backbone architecture.
  3. The experiments cover multiple datasets and models, and the ablation studies provide useful analysis of the proposed components.
Critical weaknesses

Weaknesses

  1. I feel the reported 30–50% reduction in training time per epoch is quite surprising, since the model architecture and FLOPs are unchanged. I wonder why the training can be much faster in this case? Would it be possible to also compare the total training time needed to reach the same performance?
  2. I notice that the implementation in Appendix E is a little different from the main method, especially the reward definition and PID update. I wonder why the authors choose a different implementation here? Does the theoretical analysis still apply to this setting?
  3. I feel some of the datasets are actually unimodal classification tasks. I wonder how the task-modality loss weights are defined for these tasks? Could the authors provide more details on this part?
  4. I think adaptive loss weighting has been studied in multi-task learning. Would it be possible to include some commonly used methods, such as GradNorm and FAMO? I feel this can help better understand the advantage of the proposed method.
  5. How about MoE? It's interesting to see whether the proposed method adapts to this architecture. In medical multimodal learning, MoE [1,2,3] is a popular paradigm that can potentially improve the computation efficiency, especially in multi-modal multitask learning [4]. It's worth discussing these methods in the experiment or in the related works.
  6. It would be better if there is a more detailed table for quantitative results of the figure 4.

[1] Han X, Nguyen H, Harris C, et al. Fusemoe: Mixture-of-experts transformers for fleximodal fusion[J]. Advances in Neural Information Processing Systems, 2024, 37: 67850-67900.

[2] Nguyen H, Han X, Harris C, et al. On expert estimation in hierarchical mixture of experts: Beyond softmax gating functions[J]. arXiv preprint arXiv:2410.02935, 2024.

[3] Yun S, Choi I, Peng J, et al. Flex-moe: Modeling arbitrary modality combination via the flexible mixture-of-experts[J]. Advances in Neural Information Processing Systems, 2024, 37: 98782-98805.

[4] Wu C, Shuai Z, Tang Z, et al. Dynamic modeling of patients, modalities and tasks via multi-modal multi-task mixture of experts[C]//International Conference on Learning Representations. 2025, 2025: 91267-91281.

Questions for authors

See weakness..

Reviewer_UmRSOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes BCLAN, a closed-loop loss-reweighting framework for multimodal medical learning. It combines bandit-based selection of loss-weight templates with PID-based refinement driven by validation feedback, without modifying the backbone architecture. The paper presents conditional convergence analysis and experiments across twelve medical benchmarks, reporting improved predictive performance and reduced training time.

Critical strengths

This paper proposes BCLAN, a closed-loop loss-reweighting framework for multimodal medical learning. It combines bandit-based selection of loss-weight templates with PID-based refinement driven by validation feedback, without modifying the backbone architecture. The paper presents conditional convergence analysis and experiments across twelve medical benchmarks, reporting improved predictive performance and reduced training time.

Critical weaknesses
  1. Sustained learning across tasks is insufficiently examined. Dynamically reweighting losses changes which tasks dominate training over time. Although all active tasks may continue receiving supervision, the reported results do not establish whether the controller promotes balanced, sustained learning or disproportionately favors recently emphasized tasks. Per-task learning curves and analyses of performance regression following weight shifts would help assess this concern.

  2. The generality of the allocation strategy remains unclear. The paper should examine whether a fixed controller configuration—including templates, targets, and PID gains—remains effective on unseen datasets or under distribution shifts without dataset-specific retuning.

  3. Limited methodological novelty. Both bandit optimization and PID control are established techniques. The contribution primarily lies in combining template-based loss allocation with feedback-driven refinement for medical multimodal learning. Stronger comparisons with existing adaptive loss-balancing methods and simpler stabilization mechanisms are needed to establish the distinctive benefit of this combination.

Questions for authors

The paper reports up to a 50.4% reduction in training time per epoch, while stating that the backbone architecture, parameter count, and backbone FLOPs remain unchanged. How does loss reweighting produce such a substantial reduction in per-epoch computation time? Faster convergence could reduce the number of epochs required, but does not directly explain this result. Please clarify whether sample counts, batch sizes, optimizer steps, computational paths, and timing boundaries are identical, and provide a runtime breakdown together with cumulative time to reach a common validation-performance threshold.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.