- Rating
- 2
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper proposes SurgicalMamba for online surgical phase recognition. Built on Mamba2/SSD, it introduces three main changes. First, it applies content-conditioned orthogonal transformations to recurrent states at chunk boundaries, a process called state regramming. Second, it uses boundary supervision to adjust the discretization step size and control forgetting. Third, it uses separate paths for long- and short-term modeling, each with a different state-reset policy.
Critical strengths
- The model design addresses the problems faced: the appearance of similar images at different stages and the stages needed to be forgotten correspond to the model design in the article.
- Content-conditioned state rotation is the most promising methodological contribution. Allowing intervening content to modify the orientation of stored information is an interesting direction for recurrent memory.
Critical weaknesses
-
The paper does not clearly establish its methodological advance over the backbone. Mamba2 already uses an input-dependent step size to control forgetting. The proposed intensity modulation mainly adds boundary supervision and a constrained reparameterization of that mechanism. Likewise, the dual-path design combines long- and short-term processing through different reset policies and cross-path conditioning. Besides, I feel this paper leans more toward an engineering innovation rather than a methodological one. These design choices may be useful, but the paper does not show why they are better than simpler alternatives. One important comparison would be to apply the same boundary supervision directly to Mamba2’s original step-size mechanism. Strong results on surgical phase recognition alone do not establish a methodological advance over selective state-space models.
-
The experiments do not isolate what makes state regramming effective, even though it is the paper’s main novelty claim. Removing the rotation module shows that it helps the full system. However, this does not show whether the gains depend on content-conditioned rotation subspaces and history-dependent transformations. Fixed rotations or generic learned state mixing might provide similar benefits. The norm-preservation and conditional kernel analyses establish structural properties, but they do not directly show reduced memory interference. Targeted comparisons or direct evidence of this mechanism are still needed. Without them, it is unclear whether the gains come from the proposed principle or simply from adding another architectural component. This weakens the case for accepting the paper as a general methodological contribution.
Questions for authors
The following questions concern evidence and clarification that could materially change my assessment. I am open to rerating if they can be explained properly in rebuttal.
-
Intensity modulation: Under identical boundary supervision, how does the proposed modulation compare with directly supervising Mamba2’s original selective step size or using a simple boundary-conditioned forgetting gate? If such results are unavailable, can the authors clarify what advantage of the proposed parameterization is established independently of the additional supervision?
-
State regramming: How do the proposed content- and history-conditioned rotations compare with simpler alternatives, such as fixed or input-independent rotations, content-dependent rotations in fixed planes, and nonorthogonal state mixing with the same parameter count? Which results show that the gains come from the specific rotation design, rather than simply adding a state transformation?