Review of: "Ancestral Mamba: Enhancing Selective Discriminant Space Model with Online Visual Prototype Learning for Efficient and Robust Discriminant Approach"
Luwen Huangfu · 2025
This paper presents a novel method for spatial discrete data modeling.The motivation for the method is well-de ned, and all mathematical methods are theoretically sound and proven.The method in this article can solve catastrophic forgetting and contribute to continuous learning of deep learning.The combined use of prototype learning and adaptive feedback has certain innovative contributions, which are clearly demonstrated in the ablation experiment.Experiments have demonstrated that this method can outperform many baselines and SOTA methods after multiple tasks (using both synthesized and real-world data).This article not only mentions its advantages but also covers limitations such as closed set leaning and non-task-free continual learning.This paper shows that there is a need in spatial discrete data modeling that is not being met before introducing the method that will solve the problem.In Section 3, for the instrument matrix G, it is unclear how q is selected.It is only stated that it is of order O(ln n); it would be bene cial to show how this is identi ed in practice.Additionally, guidance on how the candidate matrices in Section 4 (W_2, …, W_S) are obtained would be useful.This is discussed for the examples in Section 6 and 7; however, a more general explanation and deep dive into the logic behind the choices would help with clarity and generalization.In Section 4, both model selection and model averaging methods are well-de ned mathematically; however, there is no mention of when one should be used over the other.Should we always try model selection rst, then use model averaging if it does not work?In section 5, the mathematical proofs are well-written in showing that the estimator delta is consistent with convergence Qeios qeios.com