SAMURAI: Diverse Domain-Specific Systems

• SAMURAI is a collection of specialized systems spanning nuclear physics, computer vision, hardware security, cyberbullying detection, IoT edge AI, and quantum field computations.
• Each variant leverages state-of-the-art methods—such as high-resolution tracking, robust real-time inference, hybrid symbolic-statistical modeling, and neural field optimization—to set technical benchmarks.
• The frameworks are engineered for precise measurements, performance enhancements, and application-specific integration, supporting innovations in experimental instrumentation, adversarial detection, and 3D data retrieval.

SAMURAI encompasses a set of distinct systems and research advances across computational physics, computer vision, hardware security, cyberbullying detection, Internet-of-Things, and nuclear physics. While these share a common name, each SAMURAI is domain-specific, architected for high performance in its application, and representative of current technical trends in automation, hybrid modeling, and real-time inference.

1. Nuclear Physics and Experimental Instrumentation

In nuclear physics, SAMURAI (Superconducting Analyzer for Multi-particles from Radio Isotope Beams) denotes a superconducting dipole spectrometer at RIKEN’s Radioactive Isotope Beam Factory (RIBF), designed for high-resolution, large-acceptance tracking of rare isotopes and their decay products. With a 3 Tm field in a 1.5 m gap, it combines high-precision tracking chambers upstream and downstream of the target, achieving momentum resolution δp/p < 1×10⁻³ and angular resolution of a few milliradians (Yang et al., 2019). SAMURAI’s unique configuration—large acceptance, powerful neutron-detection arrays (NEBULA, NeuLAND), and integration with advanced trackers (STRASSE Si-barrels)—enables kinematically complete studies of multi-neutron systems and rare resonance searches.

SAMURAI supports landmark campaigns, such as studies of the tetraneutron via ^8He(p,2p)^7H→^3H+4n, using a 150 mm liquid hydrogen target (MINOS/STRASSE). Beam tracking, time-projection chambers or silicon trackers, and multi-neutron arrays provide full 4π kinematic coverage. Invariant-mass and missing-invariant-mass methods enable direct spectroscopy of few-neutron systems near threshold, with detection efficiencies of ε₄n ≈ 1% and mass resolutions ΔM₄n ≈ 1.0–1.5 MeV (Yang et al., 2019).

SAMURAI is also central to symmetry energy studies using the S$\pi$RIT TPC (Pion Reconstruction and Ion-Tracker), targeting the EoS of neutron-rich matter. The TPC’s 50.5 cm drift, 12,096-pad readout, and precise dE/dx identification provide robust pion yield ratios (π⁻/π⁺), essential for constraining the parameter γ in Eₛyₘ(ρ) = Eₛyₘ(ρ₀)(ρ/ρ₀)^γ at supra-saturation densities (Barney et al., 2020).

Advanced setups for isovector reorientation (IVR) measurements employ tensor-polarized deuteron breakup, combining PDC polarimeters, the SAMURAI magnet, and NEBULA arrays. Simulations confirm the feasibility of measuring polarization-sensitive observables with sufficient sensitivity to resolve the γ parameter to ±0.03 in a single run (Tian et al., 17 Jun 2025).

2. Visual Tracking and Computer Vision

In computer vision, SAMURAI refers to a series of methods and systems for robust, zero-shot visual object tracking, fundamentally based on the Segment Anything Model 2 (SAM2). The SAMURAI framework (Yang et al., 2024, Li et al., 28 Feb 2025, Lenhard et al., 8 Jan 2026) extends the mask-prediction backbone of SAM2 with explicit motion modeling—using a Kalman filter for motion-guided mask selection and a memory-selection mechanism informed by temporal cues.

The high-level pipeline is:

• Initialization: Given an initial bounding box (from ground truth or a detector), SAM2 predicts an object mask and derives its feature embedding.
• Framewise tracking: Each new frame is processed using the previous mask as a prompt, motion-predicted by a Kalman filter, and associated to a set of candidate masks.
• Motion-aware memory selection: Only memories where the mask score, existence score, and motion consistency exceed certain thresholds are retained for future reference.
• Association: Appearance (cosine similarity) and motion (IoU with predicted box) are fused to score candidate masks, selecting the most probable association.

Mathematically, mask selection is governed by maximizing the weighted sum

$$M_t^* = \arg\max_i \Bigl( \alpha_{kf} s_{kf,i} + (1-\alpha_{kf}) s_{mask,i} \Bigr) \quad \text{subject to } s_{obj,i} > 0.$$

where $s_{kf,i}$ is the Kalman-predicted IoU and $s_{mask,i}$ the affinity score (Yang et al., 2024).

SAMURAI achieves >7% AUC improvement on LaSOT_ext and >3% AO improvement on GOT-10k over fully supervised baselines, with robust real-time performance (20 FPS, RTX 4090) (Yang et al., 2024, Li et al., 28 Feb 2025). Identity-preserving extensions, such as integration with OSNet re-ID modules, further reduce identity switches by up to 45% (Li et al., 28 Feb 2025). Detector-augmented SAMURAI variants further fuse external detections every few frames to prevent drift and facilitate recovery after occlusions or FOV exits, delivering +0.393 success-rate improvement and up to –0.475 FNR reductions on long-duration drone surveillance sequences (Lenhard et al., 8 Jan 2026).

3. Inverse Rendering and Multimodal 3D Retrieval

SAMURAI is also the title of advanced frameworks for challenging 3D vision problems.

In inverse rendering, SAMURAI (Boss et al., 2022) performs joint optimization of shape, spatially-varying Cook–Torrance BRDF, per-view unknown illumination, and camera parameters, starting from arbitrary, unposed “in-the-wild” internet image collections. The neural field-based architecture parameterizes shape as a density field σ(x) and material as per-point Cook–Torrance components, with camera multiplexing (multi-hypothesis soft selection) and robust loss scaling to deal with weak supervision and noisy inputs.

Optimization uses a composite loss including photometric, mask, normal, BRDF, and camera regularization, e.g.,

$$L_\text{network} = \sum_j s_{p,j} \left[ \lambda_b L_\text{image} + (1-\lambda_b) L_\text{image} + L_\text{mask} + \ldots \right]$$

where $s_{p,j}$ down-weights blurry/malformed inputs (Boss et al., 2022). The method outperforms prior pose-free or latent pose methods in PSNR, SSIM, and pose error benchmarks, enabling AR/VR relightable asset extraction.

In 3D object retrieval, SAMURAI (Vo et al., 26 Jun 2025) combines CLIP-based semantic matching with shape-guided re-ranking using binary silhouette masks, plus majority voting across retrieval strategies. This pipeline achieves Recall@1 = 0.88 and MRR = 0.93 on the ROOMELSA challenge, rivaling complex end-to-end solutions. All matching operates via cosine similarity in the CLIP embedding spaces. Mask preprocessing includes connected component analysis and fixed padding. Limitations arise in multi-object masking and fragmented shapes, suggesting future work in advanced mask and spatial reasoning.

4. Hardware Security and Runtime Anomaly Detection

SAMURAI in hardware security is a runtime adversarial attack detector for AI accelerators (Rahaman et al., 10 Mar 2025). Its architecture comprises an AI Performance Counter (APC), collecting low-level layerwise metrics (e.g., sparsity $S_L$, zero counts $Z_L$, FLOPs, MACs, entropy $H_L$), and an on-chip ML analysis engine TANTO, typically an LSTM or SVM, operating entirely within accelerator boundaries.

During real-time inference, the APC vector is forwarded to TANTO, which classifies each run as benign or adversarial. Detection accuracy reaches up to 98% (LSTM) with <21% inference overhead, across models (AlexNet, ResNet, VGG) and datasets (CIFAR-10/100, MNIST). This fully hardware-based solution avoids reliance on generic CPU counters or software-instrumented adversarial detection, enabling privacy and compliance within embedded/edge AI deployments.

5. Cyberbullying Detection with Hybrid Symbolic-Statistical Methods

SAMURAI also refers to a hybrid cyberbullying detection system integrating symbolic, grammar-driven rules with compact statistical sub-models (Ptaszyński et al., 2018). The architecture features a text preparation engine (normalization, correction, adversarial obfuscation detection) and a modular detection engine where each violence phenomenon (personal attacks, threats, harassment) is handled by its own module.

Each sub-module computes a candidate phrase score $p_i$ (via a DNN or logistic regression) and applies a syntactic rule $$M_t^* = \arg\max_i \Bigl( \alpha_{kf} s_{kf,i} + (1-\alpha_{kf}) s_{mask,i} \Bigr) \quad \text{subject to } s_{obj,i} > 0.$$0 parsed from the input. Output acceptance is gated such that only phrases passing both $$M_t^* = \arg\max_i \Bigl( \alpha_{kf} s_{kf,i} + (1-\alpha_{kf}) s_{mask,i} \Bigr) \quad \text{subject to } s_{obj,i} > 0.$$1 and $$M_t^* = \arg\max_i \Bigl( \alpha_{kf} s_{kf,i} + (1-\alpha_{kf}) s_{mask,i} \Bigr) \quad \text{subject to } s_{obj,i} > 0.$$2 are flagged. This system achieves Accuracy = 0.974, Precision = 0.804, Recall = 0.843, and F1 = 0.823 on a newly annotated Formspring corpus, outperforming FastText and five commercial moderation systems (Ptaszyński et al., 2018). Error analysis attributes a significant fraction of false positives/negatives to annotation errors, highlighting the method’s consistency relative to ambiguous annotation standards.

6. IoT-Embedded ML and Edge AI

The SamurAI platform (Miro-Panades et al., 2023) is a dual-subsystem (Always-Responsive and On-Demand) IoT node combining a clock-less event-driven wake-up controller with a deep-sleep RISC-V CPU and ML accelerator (PNeuro). The AR subsystem implements a 16-bit QDI asynchronous pipeline, with sub-μW deep-sleep and 207 ns wake-up from external events. The OD subsystem includes a 64-MAC ML accelerator for up to 36 GOPS, energy efficiency of 1.3 TOPS/W, and >580 KB integrated SRAM/FeRAM.

Application scenarios demonstrate 3.5× energy savings over cloud-only processing, local inference latencies of 4.3 ms for KWS, and annual battery lifetimes on 500 mWh cells. Metrics such as Peak-to-Idle Power Ratio (FOM₁), Peak-Performance to Idle Power (FOM₂), and Versatility FOM with Retention (FOM₃) all lead state-of-the-art MCU designs, making SamurAI a prototypical solution for edge AI deployments facing intermittent sensing and complex ML workloads (Miro-Panades et al., 2023).

7. One-Loop Scattering Amplitudes in QFT

In quantitative field theory, SAMURAI stands for "Scattering AMplitudes from Unitarity-based Reduction Algorithm at the Integrand-level" (Mastrolia et al., 2010). This Fortran90 library automates the numerical evaluation of one-loop scattering amplitudes via OPP-based integrand decomposition extended to $$M_t^* = \arg\max_i \Bigl( \alpha_{kf} s_{kf,i} + (1-\alpha_{kf}) s_{mask,i} \Bigr) \quad \text{subject to } s_{obj,i} > 0.$$3-dimensional (dim. regularization) cuts. The approach uses polynomial interpolation (via Discrete Fourier Transform) to reconstruct residues at the integrand level (pentagons, boxes, triangles, bubbles, tadpoles).

For an $$M_t^* = \arg\max_i \Bigl( \alpha_{kf} s_{kf,i} + (1-\alpha_{kf}) s_{mask,i} \Bigr) \quad \text{subject to } s_{obj,i} > 0.$$4-point one-loop integral with numerator $$M_t^* = \arg\max_i \Bigl( \alpha_{kf} s_{kf,i} + (1-\alpha_{kf}) s_{mask,i} \Bigr) \quad \text{subject to } s_{obj,i} > 0.$$5 and $$M_t^* = \arg\max_i \Bigl( \alpha_{kf} s_{kf,i} + (1-\alpha_{kf}) s_{mask,i} \Bigr) \quad \text{subject to } s_{obj,i} > 0.$$6-dimensional denominators $$M_t^* = \arg\max_i \Bigl( \alpha_{kf} s_{kf,i} + (1-\alpha_{kf}) s_{mask,i} \Bigr) \quad \text{subject to } s_{obj,i} > 0.$$7, the integrand is expanded as

$$M_t^* = \arg\max_i \Bigl( \alpha_{kf} s_{kf,i} + (1-\alpha_{kf}) s_{mask,i} \Bigr) \quad \text{subject to } s_{obj,i} > 0.$$8

where residues $$M_t^* = \arg\max_i \Bigl( \alpha_{kf} s_{kf,i} + (1-\alpha_{kf}) s_{mask,i} \Bigr) \quad \text{subject to } s_{obj,i} > 0.$$9 are polynomials in $s_{kf,i}$0 and $s_{kf,i}$1 (the $s_{kf,i}$2-dimensional part). Generalized unitarity cuts and DFT-based interpolation allow for exact extraction of all coefficients, including rational terms, in a single pass.

The library interfaces with both Feynman-diagram numerators and on-shell tree amplitude products, uses external master integral libraries (QCDLoop, OneLoop), and achieves full double-precision (and higher) stability across QED/QCD benchmarks, with overhead only at problematic Gram-determinant points where it can trigger higher-precision re-evaluations. It is used in NLO calculations for LHC phenomenology (Mastrolia et al., 2010).

---

Across these domains, SAMURAI consistently refers to technically sophisticated, modular systems that prioritize interpretability, real-time operation, and performance-driven hybridization with explicit integration of physics knowledge, symbolic logic, or machine learning as appropriate. Each instantiation of SAMURAI sets a benchmark in its field for automation, accuracy, and operational robustness.