To segment medical images with distribution shifts, domain generalization (DG) has emerged as a promising setting to train models on source domains that can generalize to unseen target domains. Existing DG methods are mainly based on CNN or ViT architectures. Recently, advanced state space models, represented by Mamba, have shown promising results in various supervised medical image segmentation. The success of Mamba is primarily owing to its ability to capture long-range dependencies while keeping linear complexity with input sequence length, making it a promising alternative to CNNs and ViTs. Inspired by the success, in the paper, we explore the potential of the Mamba architecture to address distribution shifts in DG for medical image segmentation. Specifically, we propose a novel Mamba-based framework, Mamba-Sea, incorporating global-to-local sequence augmentation to improve the model's generalizability under domain shift issues. Our Mamba-Sea introduces a global augmentation mechanism designed to simulate potential variations in appearance across different sites, aiming to suppress the model's learning of domain-specific information. At the local level, we propose a sequence-wise augmentation along input sequences, which perturbs the style of tokens within random continuous sub-sequences by modeling and resampling style statistics associated with domain shifts. To our best knowledge, Mamba-Sea is the first work to explore the generalization of Mamba for medical image segmentation, providing an advanced and promising Mamba-based architecture with strong robustness to domain shifts. Remarkably, our proposed method is the first to surpass a Dice coefficient of 90% on the Prostate dataset, which exceeds previous SOTA of 88.61%. The code is available at https://github.com/orange-czh/Mamba-Sea.
Knowledge distillation is a widely used technique for compressing large language models (LLMs) by training a smaller student model to mimic a larger teacher model. Typically, both the teacher and student are Transformer-based architectures, leveraging softmax attention for sequence modeling. However, the quadratic complexity of self-attention at inference time remains a significant bottleneck, motivating the exploration of subquadratic alternatives such as structured state-space models (SSMs), linear attention, and recurrent architectures. In this work, we systematically evaluate the transferability of knowledge distillation from a Transformer teacher to nine subquadratic student architectures. Our study aims to determine which subquadratic model best aligns with the teacher's learned representations and how different architectural constraints influence the distillation process. We also investigate the impact of intelligent initialization strategies, including matrix mixing and query-key-value (QKV) copying, on the adaptation process. Our empirical results on multiple NLP benchmarks provide insights into the trade-offs between efficiency and performance, highlighting key factors for successful knowledge transfer to subquadratic architectures.
Efficiently handling long contexts in transformer-based language models with low perplexity is an active area of research. Numerous recent approaches like Linformer, Longformer, Performer, and Structured state space models (SSMs)., have not fully resolved this problem. All these models strive to reduce the quadratic time complexity of the attention mechanism while minimizing the loss in quality due to the effective compression of the long context. Inspired by the cache and virtual memory principle in computers, where in case of a cache miss, not only the needed data is retrieved from the memory, but the adjacent data is also obtained, we apply this concept to handling long contexts by dividing it into small segments. In our design, we retrieve the nearby segments in an uncompressed form when high segment-level attention occurs at the compressed level. Our en-hancements for handling long context include aggregating four attention mechanisms consisting of short sliding window attention, long compressed segmented attention, dynamically retrieving top k high attention uncompressed segments, and overlapping segments in long segment attention to avoid segment fragmentation. These enhancements result in an architecture that outperforms ex-isting SOTA architectures with an average perplexity improvement of 8.5% over similar model sizes.
Understanding bimanual hand interactions is essential for realistic 3D pose and shape reconstruction. However, existing methods struggle with occlusions, ambiguous appearances, and computational inefficiencies. To address these challenges, we propose Vision Mamba Bimanual Hand Interaction Network (VM-BHINet), introducing state space models (SSMs) into hand reconstruction to enhance interaction modeling while improving computational efficiency. The core component, Vision Mamba Interaction Feature Extraction Block (VM-IFEBlock), combines SSMs with local and global feature operations, enabling deep understanding of hand interactions. Experiments on the InterHand2.6M dataset show that VM-BHINet reduces Mean per-joint position error (MPJPE) and Mean per-vertex position error (MPVPE) by 2-3%, significantly surpassing state-of-the-art methods.
Reservoir computing (RC) represents a class of state-space models (SSMs) characterized by a fixed state transition mechanism (the reservoir) and a flexible readout layer that maps from the state space. It is a paradigm of computational dynamical systems that harnesses the transient dynamics of high-dimensional state spaces for efficient processing of temporal data. Rooted in concepts from recurrent neural networks, RC achieves exceptional computational power by decoupling the training of the dynamic reservoir from the linear readout layer, thereby circumventing the complexities of gradient-based optimization. This work presents a systematic exploration of RC, addressing its foundational properties such as the echo state property, fading memory, and reservoir capacity through the lens of dynamical systems theory. We formalize the interplay between input signals and reservoir states, demonstrating the conditions under which reservoirs exhibit stability and expressive power. Further, we delve into the computational trade-offs and robustness characteristics of RC architectures, extending the discussion to their applications in signal processing, time-series prediction, and control systems. The analysis is complemented by theoretical insights into optimization, training methodologies, and scalability, highlighting open challenges and potential directions for advancing the theoretical underpinnings of RC.
Convolutional neural networks and Transformer have made significant progresses in multi-modality medical image super-resolution. However, these methods either have a fixed receptive field for local learning or significant computational burdens for global learning, limiting the super-resolution performance. To solve this problem, State Space Models, notably Mamba, is introduced to efficiently model long-range dependencies in images with linear computational complexity. Relying on the Mamba and the fact that low-resolution images rely on global information to compensate for missing details, while high-resolution reference images need to provide more local details for accurate super-resolution, we propose a global and local Mamba network (GLMamba) for multi-modality medical image super-resolution. To be specific, our GLMamba is a two-branch network equipped with a global Mamba branch and a local Mamba branch. The global Mamba branch captures long-range relationships in low-resolution inputs, and the local Mamba branch focuses more on short-range details in high-resolution reference images. We also use the deform block to adaptively extract features of both branches to enhance the representation ability. A modulator is designed to further enhance deformable features in both global and local Mamba blocks. To fully integrate the reference image for low-resolution image super-resolution, we further develop a multi-modality feature fusion block to adaptively fuse features by considering similarities, differences, and complementary aspects between modalities. In addition, a contrastive edge loss (CELoss) is developed for sufficient enhancement of edge textures and contrast in medical images.
Despite their frequent use for change detection, both ConvNets and Vision transformers (ViT) exhibit well-known limitations, namely the former struggle to model long-range dependencies while the latter are computationally inefficient, rendering them challenging to train on large-scale datasets. Vision Mamba, an architecture based on State Space Models has emerged as an alternative addressing the aforementioned deficiencies and has been already applied to remote sensing change detection, though mostly as a feature extracting backbone. In this article the Change State Space Model is introduced, that has been specifically designed for change detection by focusing on the relevant changes between bi-temporal images, effectively filtering out irrelevant information. By concentrating solely on the changed features, the number of network parameters is reduced, enhancing significantly computational efficiency while maintaining high detection performance and robustness against input degradation. The proposed model has been evaluated via three benchmark datasets, where it outperformed ConvNets, ViTs, and Mamba-based counterparts at a fraction of their computational complexity. The implementation will be made available at https://github.com/Elman295/CSSM upon acceptance.
Hybrid LLM architectures that combine Attention and State Space Models (SSMs) achieve state-of-the-art accuracy and runtime performance. Recent work has demonstrated that applying compression and distillation to Attention-only models yields smaller, more accurate models at a fraction of the training cost. In this work, we explore the effectiveness of compressing Hybrid architectures. We introduce a novel group-aware pruning strategy that preserves the structural integrity of SSM blocks and their sequence modeling capabilities. Furthermore, we demonstrate the necessity of such SSM pruning to achieve improved accuracy and inference speed compared to traditional approaches. Our compression recipe combines SSM, FFN, embedding dimension, and layer pruning, followed by knowledge distillation-based retraining, similar to the MINITRON technique. Using this approach, we compress the Nemotron-H 8B Hybrid model down to 4B parameters with up to 40x fewer training tokens. The resulting model surpasses the accuracy of similarly-sized models while achieving 2x faster inference, significantly advancing the Pareto frontier.
This study develops and analyzes linear and nonlinear state space models for estimating the elemental composition of scrap steel used in steelmaking, with applications to Electric Arc Furnace (EAF) and Basic Oxygen Furnace (BOF) processes. The models incorporate mass balance equations and are fitted using a modified Kalman filter for linear cases and the Unscented Kalman Filter (UKF) for nonlinear cases. Using Cu and Cr as representative elements, we assess the sensitivity of model predictions to measurement noise in key process variables, including steel mass, steel composition, scrap input mass, slag mass, and iron oxide fraction in slag. Results show that the models are robust to moderate noise levels in most variables, particularly when errors are below $10\%$. However, accuracy significantly deteriorates with noise in slag mass estimation. These findings highlight the practical feasibility and limitations of applying state space models for real-time scrap composition estimation in industrial settings.
Fine-grained domain generalization (FGDG) aims to learn a fine-grained representation that can be well generalized to unseen target domains when only trained on the source domain data. Compared with generic domain generalization, FGDG is particularly challenging in that the fine-grained category can be only discerned by some subtle and tiny patterns. Such patterns are particularly fragile under the cross-domain style shifts caused by illumination, color and etc. To push this frontier, this paper presents a novel Hyperbolic State Space Hallucination (HSSH) method. It consists of two key components, namely, state space hallucination (SSH) and hyperbolic manifold consistency (HMC). SSH enriches the style diversity for the state embeddings by firstly extrapolating and then hallucinating the source images. Then, the pre- and post- style hallucinate state embeddings are projected into the hyperbolic manifold. The hyperbolic state space models the high-order statistics, and allows a better discernment of the fine-grained patterns. Finally, the hyperbolic distance is minimized, so that the impact of style variation on fine-grained patterns can be eliminated. Experiments on three FGDG benchmarks demonstrate its state-of-the-art performance.
Remote state estimation in cyber-physical systems is often vulnerable to cyber-attacks due to wireless connections between sensors and computing units. In such scenarios, adversaries compromise the system by injecting false data or blocking measurement transmissions via denial-of-service attacks, distorting sensor readings. This paper develops a Kalman filter and Rauch--Tung--Striebel (RTS) smoother for linear stochastic state-space models subject to cyber-attacked measurements. We approximate the faulty measurement model via generalized statistical linear regression (GSLR). The GSLR-based approximated measurement model is then used to develop a Kalman filter and RTS smoother for the problem. The effectiveness of the proposed algorithms under cyber-attacks is demonstrated through a simulated aircraft tracking experiment.
Events offer a novel paradigm for capturing scene dynamics via asynchronous sensing, but their inherent randomness often leads to degraded signal quality. Event signal filtering is thus essential for enhancing fidelity by reducing this internal randomness and ensuring consistent outputs across diverse acquisition conditions. Unlike traditional time series that rely on fixed temporal sampling to capture steady-state behaviors, events encode transient dynamics through polarity and event intervals, making signal modeling significantly more complex. To address this, the theoretical foundation of event generation is revisited through the lens of diffusion processes. The state and process information within events is modeled as continuous probability flux at threshold boundaries of the underlying irradiance diffusion. Building on this insight, a generative, online filtering framework called Event Density Flow Filter (EDFilter) is introduced. EDFilter estimates event correlation by reconstructing the continuous probability flux from discrete events using nonparametric kernel smoothing, and then resamples filtered events from this flux. To optimize fidelity over time, spatial and temporal kernels are employed in a time-varying optimization framework. A fast recursive solver with O(1) complexity is proposed, leveraging state-space models and lookup tables for efficient likelihood computation. Furthermore, a new real-world benchmark Rotary Event Dataset (RED) is released, offering microsecond-level ground truth irradiance for full-reference event filtering evaluation. Extensive experiments validate EDFilter's performance across tasks like event filtering, super-resolution, and direct event-based blob tracking. Significant gains in downstream applications such as SLAM and video reconstruction underscore its robustness and effectiveness.
Long-range dependencies are critical for understanding genomic structure and function, yet most conventional methods struggle with them. Widely adopted transformer-based models, while excelling at short-context tasks, are limited by the attention module's quadratic computational complexity and inability to extrapolate to sequences longer than those seen in training. In this work, we explore State Space Models (SSMs) as a promising alternative by benchmarking two SSM-inspired architectures, Caduceus and Hawk, on long-range genomics modeling tasks under conditions parallel to a 50M parameter transformer baseline. We discover that SSMs match transformer performance and exhibit impressive zero-shot extrapolation across multiple tasks, handling contexts 10 to 100 times longer than those seen during training, indicating more generalizable representations better suited for modeling the long and complex human genome. Moreover, we demonstrate that these models can efficiently process sequences of 1M tokens on a single GPU, allowing for modeling entire genomic regions at once, even in labs with limited compute. Our findings establish SSMs as efficient and scalable for long-context genomic analysis.
Cyber-Physical Systems (CPS) in domains such as manufacturing and energy distribution generate complex time series data crucial for Prognostics and Health Management (PHM). While Deep Learning (DL) methods have demonstrated strong forecasting capabilities, their adoption in industrial CPS remains limited due insufficient robustness. Existing robustness evaluations primarily focus on formal verification or adversarial perturbations, inadequately representing the complexities encountered in real-world CPS scenarios. To address this, we introduce a practical robustness definition grounded in distributional robustness, explicitly tailored to industrial CPS, and propose a systematic framework for robustness evaluation. Our framework simulates realistic disturbances, such as sensor drift, noise and irregular sampling, enabling thorough robustness analyses of forecasting models on real-world CPS datasets. The robustness definition provides a standardized score to quantify and compare model performance across diverse datasets, assisting in informed model selection and architecture design. Through extensive empirical studies evaluating prominent DL architectures (including recurrent, convolutional, attention-based, modular, and structured state-space models) we demonstrate the applicability and effectiveness of our approach. We publicly release our robustness benchmark to encourage further research and reproducibility.
In spiking neural networks (SNNs), the main unit of information processing is the neuron with an internal state. The internal state generates an output spike based on its component associated with the membrane potential. This spike is then communicated to other neurons in the network. Here, we propose a general multiple-input multiple-output (MIMO) spiking neuron model that goes beyond this traditional single-input single-output (SISO) model in the SNN literature. Our proposed framework is based on interpreting the neurons as state-space models (SSMs) with linear state evolutions and non-linear spiking activation functions. We illustrate the trade-offs among various parameters of the proposed SSM-inspired neuron model, such as the number of hidden neuron states, the number of input and output channels, including single-input multiple-output (SIMO) and multiple-input single-output (MISO) models. We show that for SNNs with a small number of neurons with large internal state spaces, significant performance gains may be obtained by increasing the number of output channels of a neuron. In particular, a network with spiking neurons with multiple-output channels may achieve the same level of accuracy with the baseline with the continuous-valued communications on the same reference network architecture.
Atmospheric turbulence is a major source of image degradation in long-range imaging systems. Although numerous deep learning-based turbulence mitigation (TM) methods have been proposed, many are slow, memory-hungry, and do not generalize well. In the spatial domain, methods based on convolutional operators have a limited receptive field, so they cannot handle a large spatial dependency required by turbulence. In the temporal domain, methods relying on self-attention can, in theory, leverage the lucky effects of turbulence, but their quadratic complexity makes it difficult to scale to many frames. Traditional recurrent aggregation methods face parallelization challenges. In this paper, we present a new TM method based on two concepts: (1) A turbulence mitigation network based on the Selective State Space Model (MambaTM). MambaTM provides a global receptive field in each layer across spatial and temporal dimensions while maintaining linear computational complexity. (2) Learned Latent Phase Distortion (LPD). LPD guides the state space model. Unlike classical Zernike-based representations of phase distortion, the new LPD map uniquely captures the actual effects of turbulence, significantly improving the model's capability to estimate degradation by reducing the ill-posedness. Our proposed method exceeds current state-of-the-art networks on various synthetic and real-world TM benchmarks with significantly faster inference speed.
We present a theoretical framework for analyzing linear attention models through matrix-valued state space models (SSMs). Our approach, Parallel Flows, provides a perspective that systematically decouples temporal dynamics from implementation constraints, enabling independent analysis of critical algorithmic components: chunking, parallelization, and information aggregation. Central to this framework is the reinterpretation of chunking procedures as computations of the flows governing system dynamics. This connection establishes a bridge to mathematical tools from rough path theory, opening the door to new insights into sequence modeling architectures. As a concrete application, we analyze DeltaNet in a generalized low-rank setting motivated by recent theoretical advances. Our methods allow us to design simple, streamlined generalizations of hardware-efficient algorithms present in the literature, and to provide completely different ones, inspired by rough paths techniques, with provably lower complexity. This dual contribution demonstrates how principled theoretical analysis can both explain existing practical methods and inspire fundamentally new computational approaches.
State-space modeling has emerged as a powerful paradigm for sequence analysis in various tasks such as natural language processing, time-series forecasting, and signal processing. In this work, we propose an \emph{Adaptive State-Space Mamba} (\textbf{ASSM}) framework for real-time sensor data anomaly detection. While state-space models have been previously employed for image processing applications (e.g., style transfer \cite{wang2024stylemamba}), our approach leverages the core idea of sequential hidden states to tackle a significantly different domain: detecting anomalies on streaming sensor data. In particular, we introduce an adaptive gating mechanism that dynamically modulates the hidden state update based on contextual and learned statistical cues. This design ensures that our model remains computationally efficient and scalable, even under rapid data arrival rates. Extensive experiments on real-world and synthetic sensor datasets demonstrate that our method achieves superior detection performance compared to existing baselines. Our approach is easily extensible to other time-series tasks that demand rapid and reliable detection capabilities.
Structured state-space models (SSMs) have emerged as a powerful architecture in machine learning and control, featuring stacked layers where each consists of a linear time-invariant (LTI) discrete-time system followed by a nonlinearity. While SSMs offer computational efficiency and excel in long-sequence predictions, their widespread adoption in applications like system identification and optimal control is hindered by the challenge of ensuring their stability and robustness properties. We introduce L2RU, a novel parametrization of SSMs that guarantees input-output stability and robustness by enforcing a prescribed L-bound for all parameter values. This design eliminates the need for complex constraints, allowing unconstrained optimization over L2RUs by using standard methods such as gradient descent. Leveraging tools from system theory and convex optimization, we derive a non-conservative parametrization of square discrete-time LTI systems with a specified L2-bound, forming the foundation of the L2RU architecture. Additionally, we enhance its performance with a bespoke initialization strategy optimized for long input sequences. Through a system identification task, we validate L2RU's superior performance, showcasing its potential in learning and control applications.
State Space Models (SSMs) are emerging as a compelling alternative to Transformers because of their consistent memory usage and high performance. Despite this, scaling up SSMs on cloud services or limited-resource devices is challenging due to their storage requirements and computational power. To overcome this, quantizing SSMs with low bit-width data formats can reduce model size and benefit from hardware acceleration. As SSMs are prone to quantization-induced errors, recent efforts have focused on optimizing a particular model or bit-width for efficiency without sacrificing performance. However, distinct bit-width configurations are essential for different scenarios, like W4A8 for boosting large-batch decoding speed, and W4A16 for enhancing generation speed in short prompt applications for a single user. To this end, we present Quamba2, compatible with W8A8, W4A8, and W4A16 for both Mamba1 and Mamba2 backbones, addressing the growing demand for SSM deployment on various platforms. Based on the channel order preserving and activation persistence of SSMs, we propose an offline approach to quantize inputs of a linear recurrence in 8-bit by sorting and clustering for input $x$, combined with a per-state-group quantization for input-dependent parameters $B$ and $C$. To ensure compute-invariance in the SSM output, we rearrange weights offline according to the clustering sequence. The experiments show that Quamba2-8B outperforms two state-of-the-art SSM quantization methods and delivers 1.3$\times$ and 3$\times$ speed-ups in the pre-filling and generation stages, respectively, while offering 4$\times$ memory reduction with only a $1.6\%$ average accuracy drop. The evaluation on MMLU shows the generalizability and robustness of our framework. The code and quantized models will be released at: https://github.com/enyac-group/Quamba.