MambONet: Efficient Mamba SSM Architectures

• MambONet is a deep learning architecture built on selective state-space (Mamba) models that achieves linear-time complexity for sequence modeling across various applications.
• Domain-specific variants, such as NetMamba for network traffic and CS-MUNet for medical imaging, demonstrate notable accuracy gains and significant speedups over traditional methods.
• The unidirectional, input-dependent parameterization and dynamic state transitions underpin MambONet's scalability and adaptability to heterogeneous and complex data.

MambONet denotes a class of deep learning architectures that leverage the unidirectional, selective state-space model—Mamba—for computationally efficient sequence modeling across disparate domains. MambONet models have been independently developed for applications including network traffic classification, graph learning, physics-informed neural operators, and visual segmentation, exploiting Mamba's linear-time complexity and dynamic parameterization to supplant traditional quadratic self-attention. The following account surveys the principal incarnations, theoretical foundations, and empirical results associated with MambONet and its derivatives (Wang et al., 2024, Pan et al., 2024, Kim et al., 2024, Mao et al., 23 Feb 2025, Chen et al., 2024).

1. Foundations: Selective State-Space Models and Unidirectional Mamba

The core of MambONet builds upon a data-dependent extension of state-space models (SSMs). Given an input sequence $x(t)$, a continuous-time SSM is defined as

$\dot{h}(t) = A h(t) + B x(t), \quad y(t) = C h(t)$

which, after zero-order-hold discretization and selective input-dependent modulation, yields the recurrent update

$$h_t = \overline{A}_t h_{t-1} + \overline{B}_t x_t, \quad y_t = C_t h_t + D_t x_t$$

with parameters $\overline{A}_t$, $\overline{B}_t$, $C_t$, $D_t$ dynamically produced by lightweight, input-driven selectors. Mamba's unidirectional, strictly forward scan ensures $O(LN E)$ time (with $L$ = sequence length, $N$ = state size, $\dot{h}(t) = A h(t) + B x(t), \quad y(t) = C h(t)$0 = expand dimension), contrasting sharply with $\dot{h}(t) = A h(t) + B x(t), \quad y(t) = C h(t)$1 for transformer attention. Key architectural variants incorporate residual nonlinearities (e.g., SiLU gating) and block-level parameterization for practical implementation (Wang et al., 2024, Pan et al., 2024).

2. Domain-Specific Architectures

2.1 Network Traffic Classification

NetMamba—termed "MambONet" in the foundational networking literature—adopts a unidirectional Mamba backbone for real-time classification of network traffic flows (Wang et al., 2024). The engineering pipeline encompasses:

• Traffic Representation: Flows are split by network 5-tuple, IP anonymized, packets cropped to fixed-size header/payload, then partitioned into non-overlapping byte "strides." Each stride is linearly embedded and appended with positional encodings and a classification token.
• Pre-training (MAE): A masked autoencoder objective masks $\dot{h}(t) = A h(t) + B x(t), \quad y(t) = C h(t)$290% of non-class tokens and reconstructs stride embeddings post-encoding, enabling robust few-shot adaptation.
• Complexity: For hidden dimension $\dot{h}(t) = A h(t) + B x(t), \quad y(t) = C h(t)$3 and small Mamba state size $\dot{h}(t) = A h(t) + B x(t), \quad y(t) = C h(t)$4, per-flow computation scales as $\dot{h}(t) = A h(t) + B x(t), \quad y(t) = C h(t)$5, outperforming transformer $\dot{h}(t) = A h(t) + B x(t), \quad y(t) = C h(t)$6.
• Results: NetMamba achieves 98.7–99.9% accuracy across 6 datacenter-scale benchmarks, with up to 60$\dot{h}(t) = A h(t) + B x(t), \quad y(t) = C h(t)$7 speedup versus transformer baselines, reduced memory usage, and strong few-shot generalization.

2.2 Heterogeneous Graph Learning

In heterogeneous node/edge graph settings, "MambaONet" (originally HeteGraph-Mamba) introduces a two-level tokenization:

• Intra-Type: For each node type, local neighborhoods (metapath instances) are encoded with per-metapath attention, aggregating into type-specific tokens.
• Inter-Type: Tokens are sorted (by metapath count, then degree) and processed sequentially with Mamba blocks per type.
• Global Update: All updated tokens are finally sequenced for a cross-type Mamba block, yielding representations for downstream (e.g., node classification) tasks.
• Empirical Findings: Outperforms 19 prior baselines on DBLP, ogbn-mag, ACM, IMDB benchmarks (e.g., F₁ = 0.9602, 0.6917), and matches the computational efficiency of scalable GNNs, with ablations confirming the necessity of both intra- and inter-type sequencing (Pan et al., 2024).

2.3 Physics Operator Learning

The "MambONet" instantiation in neural operator learning (Kim et al., 2024) targets direct mapping from potential functions $\dot{h}(t) = A h(t) + B x(t), \quad y(t) = C h(t)$8 to phase-space trajectories $\dot{h}(t) = A h(t) + B x(t), \quad y(t) = C h(t)$9 in Hamiltonian dynamics:

• Encoder: Discretized $$h_t = \overline{A}_t h_{t-1} + \overline{B}_t x_t, \quad y_t = C_t h_t + D_t x_t$$0 is passed through a stack of Mamba blocks, encoding global structural features.
• Decoder: A standard transformer block receives positional time encodings, attending over the encoder's memory to predict $$h_t = \overline{A}_t h_{t-1} + \overline{B}_t x_t, \quad y_t = C_t h_t + D_t x_t$$1 across $$h_t = \overline{A}_t h_{t-1} + \overline{B}_t x_t, \quad y_t = C_t h_t + D_t x_t$$2 query times.
• Supervised Training: The architecture is trained by MSE on trajectories generated via RK4 on randomly sampled, smooth $$h_t = \overline{A}_t h_{t-1} + \overline{B}_t x_t, \quad y_t = C_t h_t + D_t x_t$$3, avoiding error accumulation present in stepwise integrators.
• Results: Achieves MSE $$h_t = \overline{A}_t h_{t-1} + \overline{B}_t x_t, \quad y_t = C_t h_t + D_t x_t$$4, one order lower than RK4, and competitive inference latency, demonstrating the operator approach's superiority for time-global prediction and error correction.

2.4 Visual Segmentation and Medical Imaging

Vision Mamba-based "MambONet" architectures replace convolutional and self-attention modules with Mamba blocks in U-shaped encoder-decoder networks (Chen et al., 2024, Zheng et al., 20 Mar 2026):

• Patch-based VMamba Encoder: Extracted image patches are projected and processed with Visual State-Space (VSS) blocks implementing SS2D (state-space scan in four diagonal directions with $$h_t = \overline{A}_t h_{t-1} + \overline{B}_t x_t, \quad y_t = C_t h_t + D_t x_t$$5 complexity).
• Skip Connections & Decoding: Hierarchical downsample/upsample operations with skip links preserve spatial information.
• Domain-specific Priors: In CS-MUNet (Zheng et al., 20 Mar 2026), spatial Mamba blocks are modulated with pixel-wise boundary posteriors (boundary-aware state transition modulation), and channel dimensions are recast as SSM sequences (channel Mamba state aggregation) under Lipschitz constraints.
• Results: For crack segmentation, MambONet achieves up to +2.8 pp mDS and –90% FLOPs compared to CNN/transformer. For organ segmentation, CS-MUNet achieves mDice = 86.16% (UW-Madison) and 94.47% (WORD), outperforming all previous methods by +1–2.9 pp in mDice.

3. Theoretical and Computational Properties

3.1 Time Complexity and Scalability

The unidirectional Mamba SSM is characterized by strictly linear time and memory scaling in the sequence (or spatial) dimension, in contrast to quadratic costs in conventional transformers. With hardware-aware kernel fusion, inference per sequence ($$h_t = \overline{A}_t h_{t-1} + \overline{B}_t x_t, \quad y_t = C_t h_t + D_t x_t$$6 tokens) is $$h_t = \overline{A}_t h_{t-1} + \overline{B}_t x_t, \quad y_t = C_t h_t + D_t x_t$$7, where state size $$h_t = \overline{A}_t h_{t-1} + \overline{B}_t x_t, \quad y_t = C_t h_t + D_t x_t$$8 is decimated relative to hidden dim $$h_t = \overline{A}_t h_{t-1} + \overline{B}_t x_t, \quad y_t = C_t h_t + D_t x_t$$9, and $\overline{A}_t$0 in practical networking applications (Wang et al., 2024). Empirical accelerations reach up to 60$\overline{A}_t$1 for high-throughput traffic analysis, and over 15$\overline{A}_t$2 reduction in model footprint for LLM distillations (Xia et al., 20 Oct 2025).

3.2 Adaptability to Heterogeneity

MambONet demonstrates architectural flexibility for handling multiple forms of heterogeneity:

• Tokenization Heuristics: Type-wise or importance-driven ordering in graphs amplifies SSM global context propagation.
• Semantic Modulation: Input-dependent selection mechanisms adapt the state transition to local content or edge semantics.
• Domain Priors: By direct injection of boundary maps or edge priors into state-scan parameters, as exemplified by CS-MUNet, explicit domain knowledge is interleaved into the dynamical system.

4. Representative Empirical Results

Domain	Model	Best Baseline	Metric	Performance/Main Gain
Network Traffic	NetMamba	YaTC/ET-BERT	Accuracy	98.7–99.9% (+3–6 pp), 60× speed
Heterog. Graphs	MambaONet	HINormer, G-M++	F1/Acc.	0.9602/0.5763 (+1.1–5 pp), linear t.
Physics Operator	MambONet	RK4, VaRONet	MSE	$\overline{A}_t$3, order lower
Crack Segmentation	VMamba-UNet	TransUNet, CNNs	mDS/mIoU	+2.8 pp / +6.8 pp, –90% FLOPs
Medical Segmentation	CS-MUNet	Res2Net50 baselines	mDice	86.16%/94.47% (+2.9/+1.0 pp)
LLM for Networking	Mamba4Net	NetLLM	Throughput	3.96×, 5.48% params

All entries are directly traceable to respective benchmarks in the cited works (Wang et al., 2024, Pan et al., 2024, Kim et al., 2024, Chen et al., 2024, Zheng et al., 20 Mar 2026, Xia et al., 20 Oct 2025).

5. Analysis, Limitations, and Future Directions

MambONet and its derivatives enable state-space models to supersede transformers across domains where linear sequence/spatial scaling, heterogeneity, or data-dependent context propagation are required. Some limitations persist:

• Graph Domains: Two-level tokenization and metapath enumeration can become computationally expensive with growing complexity; dynamic metapath discovery remains an open problem (Pan et al., 2024, Mao et al., 23 Feb 2025).
• Operator Learning: Current MambONet operator approaches are restricted to low-dimensional or smooth function spaces; extension to high-dimensional, non-smooth or multi-modal settings needs substantial innovation (Kim et al., 2024).
• Visual Segmentation: Domain prior injection and channel sequence modeling show strong gains, but full integration of spatial, channel, and boundary priors remains a design challenge (Zheng et al., 20 Mar 2026).
• Deployment: For ultra-low latency or edge scenarios (network hardware, AR/VR), memory and parallelization optimizations are critical (Xia et al., 20 Oct 2025).

Potential future avenues include hybrid architectures (e.g., combining Mamba encoders with alternative decoders), more principled foundations for selective scheduling in SSMs, and broadening of operator learning frameworks to high-dimensional physics or medical image applications with explicit physical constraints.

---

Key References:

• (Wang et al., 2024) NetMamba: Efficient Network Traffic Classification via Pre-training Unidirectional Mamba
• (Pan et al., 2024) HeteGraph-Mamba: Heterogeneous Graph Learning via Selective State Space Model
• (Kim et al., 2024) Neural Hamilton: Can A.I. Understand Hamiltonian Mechanics?
• (Chen et al., 2024) Vision Mamba-based autonomous crack segmentation on concrete, asphalt, and masonry surfaces
• (Zheng et al., 20 Mar 2026) CS-MUNet: A Channel-Spatial Dual-Stream Mamba Network for Multi-Organ Segmentation
• (Xia et al., 20 Oct 2025) Mamba4Net: Distilled Hybrid Mamba LLMs For Networking
• (Mao et al., 23 Feb 2025) MAPN: Enhancing Heterogeneous Sparse Graph Representation by Mamba-based Asynchronous Aggregation