Designing spectral convolutional networks is a formidable task in graph learning. In traditional spectral graph neural networks (GNNs), polynomial-based methods are commonly used to design filters via the Laplacian matrix. In practical applications, however, these polynomial methods encounter inherent limitations, which primarily arise from the the low-order truncation of polynomial filters and the lack of overall modeling of the graph spectrum. This leads to poor performance of existing spectral approaches on real-world graph data, especially when the spectrum is highly concentrated or contains many numerically identical values, as they tend to apply the exact same modulation to signals with the same frequencies. To overcome these issues, in this paper, we propose Graph State Space Network (GrassNet), a novel graph neural network with theoretical support that provides a simple yet effective scheme for designing and learning arbitrary graph spectral filters. In particular, our GrassNet introduces structured state space models (SSMs) to model the correlations of graph signals at different frequencies and derives a unique rectification for each frequency in the graph spectrum. To the best of our knowledge, our work is the first to employ SSMs for the design of GNN spectral filters, and it theoretically offers greater expressive power compared with polynomial filters. Extensive experiments on nine public benchmarks reveal that GrassNet achieves superior performance in real-world graph modeling tasks.
Large Language Models (LLMs) are typically shipped with tokenizers that deterministically encode text into so-called canonical token sequences, to which the LLMs assign probability values. One common assumption is that the probability of a piece of text is the probability of its canonical token sequence. However, the tokenization of a string is not unique: e.g., the Llama2 tokenizer encodes Tokens as [Tok,ens], but [Tok,en,s] also represents the same text. In this paper, we study non-canonical tokenizations. We prove that, given a string, it is computationally hard to find the most likely tokenization for an autoregressive LLM, as well as to compute the marginal probability over all possible tokenizations. We then show how the marginal is, in most cases, indistinguishable from the canonical probability. Surprisingly, we then empirically demonstrate the existence of a significant amount of signal hidden within tokenization space. Notably, by simply aggregating the probabilities of non-canonical tokenizations, we achieve improvements across a range of LLM evaluation benchmarks for a variety of architectures, including transformers and state space models.
Recently, the state space model Mamba has demonstrated efficient long-sequence modeling capabilities, particularly for addressing long-sequence visual tasks in 3D medical imaging. However, existing generative self-supervised learning methods have not yet fully unleashed Mamba's potential for handling long-range dependencies because they overlook the inherent causal properties of state space sequences in masked modeling. To address this challenge, we propose a general-purpose pre-training framework called MambaMIM, a masked image modeling method based on a novel TOKen-Interpolation strategy (TOKI) for the selective structure state space sequence, which learns causal relationships of state space within the masked sequence. Further, MambaMIM introduces a bottom-up 3D hybrid masking strategy to maintain a masking consistency across different architectures and can be used on any single or hybrid Mamba architecture to enhance its multi-scale and long-range representation capability. We pre-train MambaMIM on a large-scale dataset of 6.8K CT scans and evaluate its performance across eight public medical segmentation benchmarks. Extensive downstream experiments reveal the feasibility and advancement of using Mamba for medical image pre-training. In particular, when we apply the MambaMIM to a customized architecture that hybridizes MedNeXt and Vision Mamba, we consistently obtain the state-of-the-art segmentation performance. The code is available at: https://github.com/FengheTan9/MambaMIM.
Neural CDEs provide a natural way to process the temporal evolution of irregular time series. The number of function evaluations (NFE) is these systems' natural analog of depth (the number of layers in traditional neural networks). It is usually regulated via solver error tolerance: lower tolerance means higher numerical precision, requiring more integration steps. However, lowering tolerances does not adequately increase the models' expressiveness. We propose a simple yet effective alternative: scaling the integration time horizon to increase NFEs and "deepen`` the model. Increasing the integration interval causes uncontrollable growth in conventional vector fields, so we also propose a way to stabilize the dynamics via Negative Feedback (NF). It ensures provable stability without constraining flexibility. It also implies robustness: we provide theoretical bounds for Neural ODE risk using Gaussian process theory. Experiments on four open datasets demonstrate that our method, DeNOTS, outperforms existing approaches~ -- ~including recent Neural RDEs and state space models,~ -- ~achieving up to $20\%$ improvement in metrics. DeNOTS combines expressiveness, stability, and robustness, enabling reliable modelling in continuous-time domains.
Deep learning based approaches are now widely used across biophysics to help automate a variety of tasks including image segmentation, feature selection, and deconvolution. However, the presence of multiple competing deep learning architectures, each with its own unique advantages and disadvantages, makes it challenging to select an architecture best suited for a specific application. As such, we present a comprehensive comparison of common models. Here, we focus on the task of segmentation assuming the typically small training dataset sizes available from biophysics experiments and compare the following four commonly used architectures: convolutional neural networks, U-Nets, vision transformers, and vision state space models. In doing so, we establish criteria for determining optimal conditions under which each model excels, thereby offering practical guidelines for researchers and practitioners in the field.
Dynamic graph learning aims to uncover evolutionary laws in real-world systems, enabling accurate social recommendation (link prediction) or early detection of cancer cells (classification). Inspired by the success of state space models, e.g., Mamba, for efficiently capturing long-term dependencies in language modeling, we propose DyG-Mamba, a new continuous state space model (SSM) for dynamic graph learning. Specifically, we first found that using inputs as control signals for SSM is not suitable for continuous-time dynamic network data with irregular sampling intervals, resulting in models being insensitive to time information and lacking generalization properties. Drawing inspiration from the Ebbinghaus forgetting curve, which suggests that memory of past events is strongly correlated with time intervals rather than specific details of the events themselves, we directly utilize irregular time spans as control signals for SSM to achieve significant robustness and generalization. Through exhaustive experiments on 12 datasets for dynamic link prediction and dynamic node classification tasks, we found that DyG-Mamba achieves state-of-the-art performance on most of the datasets, while also demonstrating significantly improved computation and memory efficiency.
The analysis of tabular data has traditionally been dominated by gradient-boosted decision trees (GBDTs), known for their proficiency with mixed categorical and numerical features. However, recent deep learning innovations are challenging this dominance. This paper investigates the use of autoregressive state-space models for tabular data and compares their performance against established benchmark models. Additionally, we explore various adaptations of these models, including different pooling strategies, feature interaction mechanisms, and bi-directional processing techniques to understand their effectiveness for tabular data. Our findings indicate that interpreting features as a sequence and processing them and their interactions through structured state-space layers can lead to significant performance improvement. This research underscores the versatility of autoregressive models in tabular data analysis, positioning them as a promising alternative that could substantially enhance deep learning capabilities in this traditionally challenging area. The source code is available at https://github.com/basf/mamba-tabular.
Recommender systems aim to estimate the dynamically changing user preferences and sequential dependencies between historical user behaviour and metadata. Although transformer-based models have proven to be effective in sequential recommendations, their state growth is proportional to the length of the sequence that is being processed, which makes them expensive in terms of memory and inference costs. Our research focused on three promising directions in sequential recommendations: enhancing speed through the use of State Space Models (SSM), as they can achieve SOTA results in the sequential recommendations domain with lower latency, memory, and inference costs, as proposed by arXiv:2403.03900 improving the quality of recommendations with Large Language Models (LLMs) via Monolithic Preference Optimization without Reference Model (ORPO); and implementing adaptive batch- and step-size algorithms to reduce costs and accelerate training processes.
Learning useful representations for continuous-time dynamic graphs (CTDGs) is challenging, due to the concurrent need to span long node interaction histories and grasp nuanced temporal details. In particular, two problems emerge: (1) Encoding longer histories requires more computational resources, making it crucial for CTDG models to maintain low computational complexity to ensure efficiency; (2) Meanwhile, more powerful models are needed to identify and select the most critical temporal information within the extended context provided by longer histories. To address these problems, we propose a CTDG representation learning model named DyGMamba, originating from the popular Mamba state space model (SSM). DyGMamba first leverages a node-level SSM to encode the sequence of historical node interactions. Another time-level SSM is then employed to exploit the temporal patterns hidden in the historical graph, where its output is used to dynamically select the critical information from the interaction history. We validate DyGMamba experimentally on the dynamic link prediction task. The results show that our model achieves state-of-the-art in most cases. DyGMamba also maintains high efficiency in terms of computational resources, making it possible to capture long temporal dependencies with a limited computation budget.
With the evolution of large language models, traditional Transformer models become computationally demanding for lengthy sequences due to the quadratic growth in computation with respect to the sequence length. Mamba, emerging as a groundbreaking architecture in the field of generative AI, demonstrates remarkable proficiency in handling elongated sequences with reduced computational and memory complexity. Nevertheless, the existing training framework of Mamba presents inefficiency with variable-length sequence inputs. Either single-sequence training results in low GPU utilization, or batched processing of variable-length sequences to a maximum length incurs considerable memory and computational overhead. To address this problem, we analyze the performance of bottleneck operators in Mamba under diverse tensor shapes and proposed PackMamba, a high-throughput Mamba that efficiently handles variable-length sequences. Diving deep into state-space models (SSMs), we modify the parallel operators to avoid passing information between individual sequences while maintaining high performance. Experimental results on an NVIDIA A100 GPU demonstrate throughput exceeding the baseline single-sequence processing scheme: 3.06x speedup on the 1.4B model and 2.62x on the 2.8B model.
In recent years, with the development of deep learning, electroencephalogram (EEG) classification networks have achieved certain progress. Transformer-based models can perform well in capturing long-term dependencies in EEG signals. However, their quadratic computational complexity poses a substantial computational challenge. Moreover, most EEG classification models are only suitable for single tasks and struggle with generalization across different tasks, particularly when faced with variations in signal length and channel count. In this paper, we introduce EEGMamba, the first universal EEG classification network to truly implement multi-task learning for EEG applications. EEGMamba seamlessly integrates the Spatio-Temporal-Adaptive (ST-Adaptive) module, bidirectional Mamba, and Mixture of Experts (MoE) into a unified framework. The proposed ST-Adaptive module performs unified feature extraction on EEG signals of different lengths and channel counts through spatial-adaptive convolution and incorporates a class token to achieve temporal-adaptability. Moreover, we design a bidirectional Mamba particularly suitable for EEG signals for further feature extraction, balancing high accuracy, fast inference speed, and efficient memory-usage in processing long EEG signals. To enhance the processing of EEG data across multiple tasks, we introduce task-aware MoE with a universal expert, effectively capturing both differences and commonalities among EEG data from different tasks. We evaluate our model on eight publicly available EEG datasets, and the experimental results demonstrate its superior performance in four types of tasks: seizure detection, emotion recognition, sleep stage classification, and motor imagery. The code is set to be released soon.
Current state-of-the-art (SOTA) methods in 3D Human Pose Estimation (HPE) are primarily based on Transformers. However, existing Transformer-based 3D HPE backbones often encounter a trade-off between accuracy and computational efficiency. To resolve the above dilemma, in this work, we leverage recent advances in state space models and utilize Mamba for high-quality and efficient long-range modeling. Nonetheless, Mamba still faces challenges in precisely exploiting local dependencies between joints. To address these issues, we propose a new attention-free hybrid spatiotemporal architecture named Hybrid Mamba-GCN (Pose Magic). This architecture introduces local enhancement with GCN by capturing relationships between neighboring joints, thus producing new representations to complement Mamba's outputs. By adaptively fusing representations from Mamba and GCN, Pose Magic demonstrates superior capability in learning the underlying 3D structure. To meet the requirements of real-time inference, we also provide a fully causal version. Extensive experiments show that Pose Magic achieves new SOTA results ($\downarrow 0.9 mm$) while saving $74.1\%$ FLOPs. In addition, Pose Magic exhibits optimal motion consistency and the ability to generalize to unseen sequence lengths.
This paper examines the mathematical foundations of transformer architectures, highlighting their limitations particularly in handling long sequences. We explore prerequisite models such as Mamba, Vision Mamba (ViM), and LV-ViT that pave the way for our proposed architecture, DeMansia. DeMansia integrates state space models with token labeling techniques to enhance performance in image classification tasks, efficiently addressing the computational challenges posed by traditional transformers. The architecture, benchmark, and comparisons with contemporary models demonstrate DeMansia's effectiveness. The implementation of this paper is available on GitHub at https://github.com/catalpaaa/DeMansia
As one of the most representative DL techniques, Transformer architecture has empowered numerous advanced models, especially the large language models (LLMs) that comprise billions of parameters, becoming a cornerstone in deep learning. Despite the impressive achievements, Transformers still face inherent limitations, particularly the time-consuming inference resulting from the quadratic computation complexity of attention calculation. Recently, a novel architecture named Mamba, drawing inspiration from classical state space models (SSMs), has emerged as a promising alternative for building foundation models, delivering comparable modeling abilities to Transformers while preserving near-linear scalability concerning sequence length. This has sparked an increasing number of studies actively exploring Mamba's potential to achieve impressive performance across diverse domains. Given such rapid evolution, there is a critical need for a systematic review that consolidates existing Mamba-empowered models, offering a comprehensive understanding of this emerging model architecture. In this survey, we therefore conduct an in-depth investigation of recent Mamba-associated studies, covering three main aspects: the advancements of Mamba-based models, the techniques of adapting Mamba to diverse data, and the applications where Mamba can excel. Specifically, we first review the foundational knowledge of various representative deep learning models and the details of Mamba-1&2 as preliminaries. Then, to showcase the significance of Mamba for AI, we comprehensively review the related studies focusing on Mamba models' architecture design, data adaptability, and applications. Finally, we present a discussion of current limitations and explore various promising research directions to provide deeper insights for future investigations.
Ultra-high-definition (UHD) technology has attracted widespread attention due to its exceptional visual quality, but it also poses new challenges for low-light image enhancement (LLIE) techniques. UHD images inherently possess high computational complexity, leading existing UHD LLIE methods to employ high-magnification downsampling to reduce computational costs, which in turn results in information loss. The wavelet transform not only allows downsampling without loss of information, but also separates the image content from the noise. It enables state space models (SSMs) to avoid being affected by noise when modeling long sequences, thus making full use of the long-sequence modeling capability of SSMs. On this basis, we propose Wave-Mamba, a novel approach based on two pivotal insights derived from the wavelet domain: 1) most of the content information of an image exists in the low-frequency component, less in the high-frequency component. 2) The high-frequency component exerts a minimal influence on the outcomes of low-light enhancement. Specifically, to efficiently model global content information on UHD images, we proposed a low-frequency state space block (LFSSBlock) by improving SSMs to focus on restoring the information of low-frequency sub-bands. Moreover, we propose a high-frequency enhance block (HFEBlock) for high-frequency sub-band information, which uses the enhanced low-frequency information to correct the high-frequency information and effectively restore the correct high-frequency details. Through comprehensive evaluation, our method has demonstrated superior performance, significantly outshining current leading techniques while maintaining a more streamlined architecture. The code is available at https://github.com/AlexZou14/Wave-Mamba.
Remote Photoplethysmography (rPPG) enables non-contact physiological signal extraction from facial videos, offering applications in psychological state analysis, medical assistance, and anti-face spoofing. However, challenges such as motion artifacts, lighting variations, and noise limit its real-world applicability. To address these issues, we propose PhysMamba, a novel dual-pathway time-frequency interaction model based on Synergistic State Space Duality (SSSD), which for the first time integrates state space models with attention mechanisms in a dual-branch framework. Combined with a Multi-Scale Query (MQ) mechanism, PhysMamba achieves efficient information exchange and enhanced feature representation, ensuring robustness under noisy and dynamic conditions. Experiments on PURE, UBFC-rPPG, and MMPD datasets demonstrate that PhysMamba outperforms state-of-the-art methods, offering superior accuracy and generalization. This work lays a strong foundation for practical applications in non-contact health monitoring, including real-time remote patient care.
Structured State Space Models (SSMs) have emerged as compelling alternatives to Transformer architectures, offering linear-time complexity and superior performance in various sequence modeling tasks. Despite their advantages, SSMs like the original Mamba-2 face training difficulties due to the sensitivities introduced by the extended series of recurrent matrix multiplications. In this paper, we propose an advanced architecture that mitigates these challenges by decomposing A-multiplications into multiple groups and optimizing positional encoding through Grouped Finite Impulse Response (FIR) filtering. This new structure, denoted as Grouped FIR-enhanced SSM (GFSSM), employs semiseparable matrices for efficient computation. Furthermore, inspired by the "attention sink" phenomenon identified in streaming language models, we incorporate a similar mechanism to enhance the stability and performance of our model over extended sequences. Our approach further bridges the gap between SSMs and Transformer architectures, offering a viable path forward for scalable and high-performing sequence modeling.
Transformers have become the de-facto standard model in artificial intelligence since 2017 despite numerous shortcomings ranging from energy inefficiency to hallucinations. Research has made a lot of progress in improving elements of transformers, and, more generally, deep learning manifesting in many proposals for architectures, layers, optimization objectives, and optimization techniques. For researchers it is difficult to keep track of such developments on a broader level. We provide a comprehensive overview of the many important, recent works in these areas to those who already have a basic understanding of deep learning. Our focus differs from other works, as we target specifically novel, alternative potentially disruptive approaches to transformers as well as successful ideas of recent deep learning. We hope that such a holistic and unified treatment of influential, recent works and novel ideas helps researchers to form new connections between diverse areas of deep learning. We identify and discuss multiple patterns that summarize the key strategies for successful innovations over the last decade as well as works that can be seen as rising stars. Especially, we discuss attempts on how to improve on transformers covering (partially) proven methods such as state space models but also including far-out ideas in deep learning that seem promising despite not achieving state-of-the-art results. We also cover a discussion on recent state-of-the-art models such as OpenAI's GPT series and Meta's LLama models and, Google's Gemini model family.
The outdoor vision systems are frequently contaminated by rain streaks and raindrops, which significantly degenerate the performance of visual tasks and multimedia applications. The nature of videos exhibits redundant temporal cues for rain removal with higher stability. Traditional video deraining methods heavily rely on optical flow estimation and kernel-based manners, which have a limited receptive field. Yet, transformer architectures, while enabling long-term dependencies, bring about a significant increase in computational complexity. Recently, the linear-complexity operator of the state space models (SSMs) has contrarily facilitated efficient long-term temporal modeling, which is crucial for rain streaks and raindrops removal in videos. Unexpectedly, its uni-dimensional sequential process on videos destroys the local correlations across the spatio-temporal dimension by distancing adjacent pixels. To address this, we present an improved SSMs-based video deraining network (RainMamba) with a novel Hilbert scanning mechanism to better capture sequence-level local information. We also introduce a difference-guided dynamic contrastive locality learning strategy to enhance the patch-level self-similarity learning ability of the proposed network. Extensive experiments on four synthesized video deraining datasets and real-world rainy videos demonstrate the effectiveness and efficiency of our network in the removal of rain streaks and raindrops. Our code and results are available at https://github.com/TonyHongtaoWu/RainMamba.
This paper describes a discrete state-space model -- and accompanying methods -- for generative modelling. This model generalises partially observed Markov decision processes to include paths as latent variables, rendering it suitable for active inference and learning in a dynamic setting. Specifically, we consider deep or hierarchical forms using the renormalisation group. The ensuing renormalising generative models (RGM) can be regarded as discrete homologues of deep convolutional neural networks or continuous state-space models in generalised coordinates of motion. By construction, these scale-invariant models can be used to learn compositionality over space and time, furnishing models of paths or orbits; i.e., events of increasing temporal depth and itinerancy. This technical note illustrates the automatic discovery, learning and deployment of RGMs using a series of applications. We start with image classification and then consider the compression and generation of movies and music. Finally, we apply the same variational principles to the learning of Atari-like games.