Annealed Multiple Choice Learning: Overcoming limitations of Winner-takes-all with annealing

We introduce Annealed Multiple Choice Learning (aMCL) which combines simulated annealing with MCL. MCL is a learning framework handling ambiguous tasks by predicting a small set of plausible hypotheses. These hypotheses are trained using the Winner-takes-all (WTA) scheme, which promotes the diversity of the predictions. However, this scheme may converge toward an arbitrarily suboptimal local minimum, due to the greedy nature of WTA. We overcome this limitation using annealing, which enhances the exploration of the hypothesis space during training. We leverage insights from statistical physics and information theory to provide a detailed description of the model training trajectory. Additionally, we validate our algorithm by extensive experiments on synthetic datasets, on the standard UCI benchmark, and on speech separation.

Paper

Similar papers

Peer review

Reviewer qUke7/10 · confidence 2/52024-07-10

Summary

The paper introduces Annealed Multiple Choice Learning (aMCL), a method that integrates simulated annealing with Multiple Choice Learning (MCL), in applications where the output label may be ambiguous, and many values may be plausible given the same input. The authors show that problems arise with the use of winner-takes-all in the optimization step of MCL, where the output space is partitioned in Voronoi cells based on the individual predictors. Instead, the authors suggest to soft-weight the predictors to replace hard-WTA assignments, in a way that depends on a temperature parameter. Low values of the temperature make the weighting scheme behave close to hard-WTA (sharpening the softmin distribution of weights).

Strengths

- The integration of simulated annealing and MCL is novel and effective; I also like that the objective in eq. (6) becomes fully differentiable, without any non-differentiable assignment or operators. - Extensive theoretical analysis. - Extensive validation over a large number of datasets, although from the results in seems that epsilon-MCL seems to perform better, overall.

Weaknesses

- The introduction of the temperature schedule as hyperparameter complicates the use of the method, while it seems that epsilon-MCL would be both easier (one hyperparameter, but fixed, without schedule) and overall better performing. - I am not sure how representative the UCI datasets are of real world problems, but that is not a big problem. - Minor: typo? 'broaden impact' -> 'broadeR impact'?

Questions

N/A

Rating

7

Confidence

2

Soundness

3

Presentation

2

Contribution

3

Limitations

- Since the authors argue that WTA is sensitive in particular to initialization, it would have been interesting to see experiments comparing the robustness of vanilla MCL vs aMCL over different initializations.

Reviewer WQVX5/10 · confidence 3/52024-07-11

Summary

I read the response and found it convincing, and I appreciate the new results. I decided to raise my score. -- This paper aims to tackle the local minima problems in MCL optimization. Inspired by simulated annealing, a temperature-controlled soft assignment is used and then soft-cells are directly optimized in the MCL process. Some theory and 2d analysis sheds insight into why this approach works. This work seems well motivated and reasonable, but the experimental gains on real datasets are quite small. Additionally, the performance seems similar to the older epsilon-MCL approach, in which some gradient is used to increase the weighting on non-selected cells. notes from reading the paper: -Multiple choice learning handles ambiguous tasks by producing a set of hypotheses. -Hypotheses trained using winner-take-all technique, to encourage diversity. -This paper aims to apply simulated annealing ideas to MCL. -Try to increase variance in the annealing process, to address MCL's issue of falling into local minima based on the initialization. -aMCL uses softmin assignment, followed by gradient steps on the soft cells.

Strengths

-The basic theoretical analysis is nice, providing relevant bounds on the performance of the algorithm. -The analytical results also nicely complement the theory.

Weaknesses

-The empirical results seem lackluster. Epsilon-MCL and aMCL seem to work roughly equally well on the UCI datasets. For the speech separation datasets, the improvement over MCL seems quite small, and Epsilon-MCL doesn't have results presented.

Questions

-Because it could be difficult to tune the temperature schedule over the course of training, I wonder if it could be possible to move the temperature schedule into the inference process (a bit like what's done in diffusion generative models)? I.e., we would train a few models with different temperature levels, and force the cell assignment to be consistent with the higher temperature level.

Rating

5

Confidence

3

Soundness

3

Presentation

3

Contribution

2

Limitations

Limitations aren't discussed in much detail in the paper.

Reviewer WAaS7/10 · confidence 3/52024-07-19

Summary

The paper proposes to apply deterministic simulated annealing to multiple choice learning (MCL) as a means to mitigate some of the drawbacks associated with the winner-takes-all (WTA) scheme used to train MCLs, such as sensitivity to initialization and hypothesis collapse. They demonstrate that the proposed annealed CML (aCML) method works well in practice, on par with previous approaches, and allows for an interesting and sound theoretical analysis of the training trajectory.

Strengths

- The proposed method, aMCL, is very well motivated both theoretically and based on existing open research questions in multiple choice learning. Moreover, the mathematical development of the paper is very clear and seems sound. - The paper is very well written and easy to follow.

Weaknesses

- The main weakness of the paper is that the experimental results, with the exception of the quite enlightening toy experiments, are somewhat underwhelming and aCML does not seem to outperform the baselines in most cases. I think the paper could benefit from a more in-depth discussion as to why that is the case. - Admittedly, I am not an expert on the CML literature and maybe guilty of hindsight bias, but the idea of applying simulated annealing to CML seems quite intuitive and, one might argue, somewhat incremental. For instance, simply annealing $\epsilon$ in $\epsilon$-CML (Rupprecht et al., 2017) seems quite natural, and perhaps a good baseline for the authors to compare against. That does not take away from the solid mathematical motivation and theoretical results in the paper though. ### Minor issues - Line 201: “in no longer” should probably be “is no longer”. - The work of Rupprecht et al. is mentioned under different names (Relaxed WTA, $\epsilon$-WTA and $\epsilon$-CML) which can be a bit confusing. - It is not entirely clear to me what the authors want to show in Figure 2. - The robustness result in Figure 8 is quite interesting and could be expanded upon (perhaps with more random seeds) or at least mentioned in the main paper. ### References Christian Rupprecht, Iro Laina, Robert DiPietro, Maximilian Baust, Federico Tombari, Nassir Navab, and Gregory D Hager. Learning in an uncertain world: Representing ambiguity through multiple hypotheses. In Proceedings of the IEEE international conference on computer vision, pages 3591–3600, 2017.

Questions

1. Do the authors compare against the score-based method of Letzelter et al.? In line 232, it is suggested that this was one of the baselines, but it was not mentioned again in the paper, unless I missed it. 2. On a similar note, could the authors elaborate on the relation between aMCL and rMCL (Letzelter et al., 2023)? It seems to be both have similar objectives, but the assignments in aCML are a function of the temperature of the system and the loss function, while in rMCL the assignments are learnable. If that is correct, would be fair to say that, given enough data and sufficient learnable parameters, we can always expect rMCL to outperform aCML? 3. It is somewhat surprising that aCML does not outperform $\epsilon$-CML in many cases. Do the authors have any intuition as to why that could be? Is it because the temperature schedule is hard to tune or aCML would require a larger number of epochs to fully converge? Or maybe the datasets are not particularly sensitive to bias of $\epsilon$-CML? 4. Have the authors considered stochastic simulated annealing as well? ### References Letzelter, Victor, et al. "Resilient Multiple Choice Learning: A learned scoring scheme with application to audio scene analysis." Advances in neural information processing systems 36 (2023).

Rating

7

Confidence

3

Soundness

4

Presentation

3

Contribution

3

Limitations

The limitations section is very well written and covers all possible shortcomings of the model and analysis proposed in the paper that I could think of.

Reviewer WAaS2024-08-12

I truly appreciate the detailed answers about the method and related literature. All my questions in that regard were completely satisfied, and I encourage the authors to include these clarifications in the final version of the paper. Unfortunately, I still think the empirical results are somewhat underwhelming, but I see now how aCML could strike a valuable balance between distortion and RMSE. I think the authors should make this trade-off more clear, instead of simply showing Table 1 with RMSE results in the main paper. Figure B in the extra results is particularly enlightening, and I'd argue the results for the UCI dataset should also highlight this trade-off between distortion and RMSE. Looking at Tables 1 and 3, I can see aCML results are usually in-between $\epsilon$-MCL and MCL, but I think this should be made evident and not require comparing two tables in very different parts of the paper. All in all, I am happy to raise my score to 7 provided these extra clarifications and experimental results are added to the paper.

Authorsrebuttal2024-08-12

We thank the reviewer for their positive feedback. In the next version of the paper, we will include clarifications about the method and the related literature, as well as the additional experimental results exhibiting a tradeoff between distortion and RMSE, which will be emphasized in the main paper.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC