Edge0
Device-native foundation models: 145B on a laptop, 35B on a phone. Latent MoE, SSD-streamed weights, a router that predicts loads a step ahead. Mac, iPhone, Android, Windows.
12 repos · 0 papers · 47 scholar
Showing 59 of 59 items
Device-native foundation models: 145B on a laptop, 35B on a phone. Latent MoE, SSD-streamed weights, a router that predicts loads a step ahead. Mac, iPhone, Android, Windows.
Powered by NVIDIA cuOpt: a GPU-accelerated portfolio optimization toolkit for building, backtesting, and scaling Mean-CVaR and Mean-Variance investment workflows with CUDA-X Data Science.
LabVIEW Development Environment for Electronic Circuit Design and Process Automation
Emerging Threats and Countermeasures in Neuromorphic Systems: A Survey. By designing and analyzing Memristor Devices for Neuromorphic Computing, Spiking Neural Networks (SNNs), Physically Unclonable Functions (PUFs), True Random Number Generators (TRNGs), we are investigating their hardware and software security (attacks and defenses).
Synthetic Biology Meets Neuromorphic Computing: Towards a Bio-Inspired Olfactory Perception System
Knowledge base on the brain: anatomy, function, and brain-inspired learning models
Pandrosion root geometry: V22 paper, Lean proofs, interactive geometric constructions and analog SPICE prototypes.
Ava: a brain-like language model in pure Rust. Spiking neurons, fast synapses, sparse codes, and sleep with synaptic consolidation. No position embeddings, almost no attention. Tiny today, built to scale. Source-available (PolyForm Noncommercial).
We implemented the research paper on CICAM operation named "CICAM: A Reconfigurable SRAM-Based Computing-in-Content-Addressable Memory Architecture Using Charge Sharing" on Cadence Virtuoso at 28nm node CMOS Technology.
A website for spatial optical computing and intelligence group in Photonics Intelligence Lab.
Physics-aware device and crossbar simulation for nonlinear analog neural computing. Code for the LASCAS 2027 paper.
Silicon Photonic Neural Network (PNN): First-principles Ansys Lumerical (FDE & INTERCONNECT) co-design with Hardware-Aware Trained (HAT) PyTorch optical computing engine.
End-to-end VITS text-to-speech models synthesize natural speech in a single forward pass, yet standard ONNX exports execute monolithically over an entire clause, delaying Time-To-First-Audio (TTFA) to several seconds and spiking activation memory. Client-side deployment across Indic scripts faces four further bottlenecks: 138.8 MB float32 checkpoints per language strain browser caches, mobile WebGPU stalls for 6-14 s compiling convolution shaders while still running slower than real time ($1.44\times$ RTF), small WebAssembly pthread pools stall at barriers on big-LITTLE mobile CPUs, and mobile Web Audio pipelines drop frames or mute output under AudioWorklet contention and iOS WKWebView rules. We present Ninada, an open-source 24 kHz TTS system covering English, Classical Sanskrit, and ten Indian languages (sa, en, hi, kn, mr, bn, ta, te, ml, gu, pa, or) across a capability-routed 48-checkpoint fleet: 24 studio HiFi-GAN v1 specialists (27.9 MB W8A32+FP16 ONNX) for desktop/WebGPU and 24 distilled c8c8_384 iSTFTNet specialists (21.7 MB ONNX, $2.64\times$ fewer vocoder MACs via an exact NOLA dual-window ConvTranspose1d inverse STFT with zero complex ops) for mobile WASM, plus a 0.0 MB zero-dependency Klatt fallback. At runtime, Ninada splits a single ONNX file in memory in $<15$ ms into prior, flow, and windowed vocoder sub-sessions streaming 12-frame halos with raised-cosine crossfades. On $N = 480$ held-out 24 kHz studio utterances, Ninada achieves 3.22 dB 0-8 kHz MCD (3.09 dB mobile iSTFTNet; 1.33 dB 24-voice distillation MCD, vs. 5.17 dB for the 1,526.8 MB 16 kHz Meta MMS fleet) and 6.07 dB 8-12 kHz error, delivering 89-155 ms audible startup on desktop WebGPU and 538-649 ms zero-underrun streaming at 0.75-0.76x RTF on a physical Google Pixel 10 Pro XL.
The sampling-to-update delay and fixed-point word length that a PV-STATCOM current controller can tolerate during ride-through are rarely quantified against ride-through metrics, and complete fixed-point FPGA fast paths are rarely reported with post-route timing. A dq-domain simulation reproduces an exact discrete-time model of the current loop to within 1.2 × 10−6 pu. A stationary-frame co-simulation that runs the bit-accurate fast path, and in 12 cases the VHDL itself in lock step, repeats a 640-case balanced-sag matrix: no case fails at short-circuit ratio (SCR) of 6 or more with 0.5 pu or more retained source voltage (0 of 192), and 116 of 128 fail at SCR of 3 or less with 0.2 pu or less. Selected cases are re-run with dead time, sensor errors, jitter, inductor saturation and single- and two-phase-to-ground faults. Current-step overshoot rises from 0% to 82% as the effective delay grows from 50 to 350 µs. With the full core, 16 fractional bits reproduce the floating-point pass/fail outcome in all 64 averaged-plant cases. A Q12.20 VHDL fast path matches its integer model bit for bit and meets 100 MHz out of context (worst slack +0.219 ns; 6480 LUTs, 65 DSP48E1); the control tick takes 208 clocks (2.08 µs).
Implementation of minimum order low-pass FIR filter for communication systems applications is discussed in this paper. The FIR filter is implemented based on transposed architecture and implemented on two FPGAs from Xilinx Inc, which include Spartan-3E (xc3s500efg320-5) and Virtex-2P (xc2vp30ff1152-7). VHDL language is considered for implementation of FIR filter architecture. The primary focus of this work is comparative analysis of these two FPGAs in terms of hardware utilization and performance for low-pass FIR filter implementation. In this work, Equiripple and Kaiser windowing approaches are used for designing a low-pass filter. It is observed that Equiripple approach leads to more effective design since it fulfills the necessary design requirements with fewer taps when compared to the Kaiser Window approach. Moreover, FCSD approach is used for representing filter coefficients. FCSD approach reduces hardware complexity, hardware area consumption and propagation delay. Simulation and synthesis results suggest that the Equiripple FIR filter implemented on Virtex-2P FPGA runs 27.48% faster than the corresponding FIR filter implemented on Spartan-3E FPGA device.
The finite-field multiplication is the most computationally intensive operation in the binary-field Elliptic Curve Cryptography (ECC) processors, affecting significantly the hardware area, delay, and overall performance. We propose an optimized overlap-free 5-way Karatsuba multiplier for polynomial multiplication over GF (2233) in the context of FPGA-based ECC hardware accelerators. The proposed architecture employs 5-way recursive decomposition and an overlap-free recombination technique to minimize the number of redundant XOR operations, reduce hardware complexity and improve lookup table (LUT) utilization. To achieve a balanced trade-off between resource utilization and critical-path delay, we adopt a hybrid implementation strategy that preserves the sub-quadratic computational complexity of the Karatsuba algorithm. The proposed multiplier was implemented in Verilog HDL and synthesized on Xilinx Virtex-5 FPGA platform. Experimental results show that the proposed architecture uses 19,240 LUTs and achieves a maximum operating frequency of 79.1 MHz with the critical-path delay reduced to 12.63 ns. The proposed design results in 14.6% delay reduction, 222 LUTs reduction in LUT utilization and 15.6% improvement in area–delay product over the conventional 5-way Karatsuba architecture. These results show that the proposed optimized overlap-free 5-way Karatsuba multiplier can be used as an efficient and scalable solution for high-performance GF((2233) ECC processors for embedded and Internet-of-Things security applications.
FPGA design flows frequently requires multiple rounds of synthesis, verification, fault detection, and redesign, which can take a lot of time and require human input. Most current verification tools only look for structural errors in netlists and don’t have ways to fix them automatically. This paper proposes a Graph Neural Network (GNN) based learning framework for the automatic reconstruction of defective FPGA netlists. The synthesized netlist is transformed into an instance–net bipartite graph, wherein faults are emulated by severing instance–net connections. A two-stage learning method is used: (i) a transformer-based embedding model captures the structural and relational features of nodes, (ii) Equivariant Graph Neural Networks (EGNN) model performs edge prediction for circuit completion, identify and reconstruct missing or incorrect connections. Synthesized results of full-adder circuit designs 1-bit full adder 4-bit full adder, 8-bit full adder, and 16-bit full adder show that it is possible to recover missing edges and restore structural connectivity. The expected full adder topology is compared with reconstructed netlist to validate the effectiveness of reconstruction.
Accurate electro-optic co-simulation is essential for electronic-photonic design automation and high-speed electronic-photonic integrated circuit design. However, existing EDA-compatible compact models still provide limited accuracy for >25-GHz traveling-wave Mach-Zehnder modulators, mainly because microwave loss, impedance/velocity mismatch, and bias-dependent carrier-depletion effects are not fully captured. In this paper, an EDA-compatible distributed behavioral model is proposed for silicon carrier-depletion traveling-wave Mach-Zehnder modulators. The optical modulation behavior is implemented in Verilog-A, while the traveling-wave electrodes are modeled using segmented equivalent circuits with distributed parameters extracted from 3-D electromagnetic S-parameters. By incorporating electrode propagation loss and bias-dependent PN-junction capacitance, the proposed model enables accurate electro-optic simulation in both frequency and time domains within Cadence Virtuoso. The model accurately predicts the bias-dependent electro-optic 3-dB bandwidth roll-off in frequency-domain simulation. In time-domain verification, a 50-Gb/s transient simulation produces a clear optical eye diagram with a 4.1-dB extinction ratio, showing less than 0.1-dB deviation from the reference result. This agreement confirms the fidelity of the proposed model.
Reversible computing paradigm has found a very practical application to the elimination of the loss of information in the digital system to allow exceptionally low power dissipation. This paper will propose a proposal of a high-performance 32-bit Arithmetic Logic Unit (ALU) that consists of reversible logic gates, flip-flop clock gating, and an Adaptive Quantum Cost Reversible ALU design to produce a high-performing system with high power efficiency. There is also a proposed architecture that has adaptive quantum cost dynamic optimization and clever clock gating as opposed to the traditional reversible ALUs in which the gate configurations were fixed. The design facilitates 16 reversible arithmetic and logical operations with the help of Peres, Fredkin, Feynman, Double Feynman and Toffoli gates. When coded on Xilinx Artix-7 FPGAs with Verilog HDL, the proposed implementation achieves major gains, such as energy savings of 95 per cent, reduced delay of 48 per cent, quantum cost of 288, and garbage output is 87-per cent lower, and therefore sets a new standard in reversible computing energy efficiency.
This work explores the capability of 8T-2MTJ bitcell, originally investigated for ternary search operation, to support both associative search and In-Memory Computing (IMC) within the same Non-Volatile Ternary Content-Addressable Memory (NV-TCAM) framework. The proposed approach reuses the existing search-line control mechanism of the bitcell and enables reconfigurable T CAM/IMC o peration w ithout m odifying the core cell structure. By appropriately configuring the search-line inputs, the 8T-2MTJ bitcell realizes AND/NAND, OR/NOR, and XOR/XNOR operations while preserving conventional TCAM functionality, including logic ‘0’, logic ‘1’, and ‘don't care’ states. A 4 × 4 array is implemented with write, wordline, and configurable IMC/TCAM mode d rivers, and the bitcell layout is designed in 18 nm FinFET technology with an area of $0.694 \mu \mathrm{m}^{2}$. Circuit-level simulations using a Magnetic Tunnel Junction (MTJ) Verilog-A compact model validate correct write, search, and logic operations. Process-corner and Monte Carlo analyses further demonstrate robust functionality under variation, with sub-nanosecond mean logic delays. These results show that the proposed reconfigurable v oltage-divider-based 8 T-2MTJ bitcell improves hardware utilization by integrating search and logic computation in a compact non-volatile memory platform.
Trusted Platform Modules (TPMs) serve as hardware roots of trust in secure System on Chip (SoC) environments, supporting essential security services such as secure key exchange, digital signatures, authentication, and cryptographic key management. Standalone validation of cryptographic hardware accelerators is an essential step for ensuring functional correctness before their integration into secure SoC platforms. Existing SPI and AXI based interface architectures primarily focus on communication and processor integration, with limited support for reusable pre-integration validation of cryptographic IP cores. This paper presents a SPI to AXI hardware validation framework for standalone verification of AXI enabled AES accelerators. The proposed architecture employs a structured packet based communication mechanism with dual clock domain synchronization and FIFO based buffering to enable reliable data transfer between the SPI interface and the AXI based AES accelerator. The framework supports AES encryption and decryption for all standard modes of operation with variable key sizes. The design is functionally verified using NIST SP 800-38A reference test vectors, and validated on a Xilinx Virtex UltraScale VCU108 FPGA, achieving an operating frequency of 333.33 MHz, utilizing 925 slices, and delivering a throughput of 46.9 Gbps. ASIC synthesis using the SCL 180 nm further demonstrates the feasibility of the proposed architecture for hardware realization. The proposed framework provides a validation platform that enables independent verification of AXI based cryptographic accelerators before their integration into the RISC-V based BHARAT-TPM.
The increasing demand for energy efficient and high performance digital systems necessitates advanced power optimization methods in modern VLSI design. This work presents a machine learning based driven adaptive clock gating framework for RTL-based digital systems, implemented through a Python integrated co-simulation environment. A 12-bit ADC is modeled in Verilog HDL and switching activity data is utilized to train multiple deep learning architectures, including CNN, Seq2Seq, RNN, LSTM, and Transformer, for dynamic prediction of clock enable signals. The generated dataset incorporates cycle level behavior, output transitions, and temporal dependencies to enable intelligent gating decisions. Experimental results indicate that the Convolutional Neural Network (CNN) achieves the highest performance, with approximately 90.0% accuracy and the lowest loss, followed by Seq2Seq (88.08%), RNN (87.69%), and LSTM (85.00%), while the Transformer model shows comparatively lower accuracy (55.38%) due to its high data dependency, CNN is preferred over other models due to its superior feature extraction capability, faster convergence, moderate computational complexity, and strong suitability for small-to-medium datasets, making it highly effective for RTL-level integration. In Waveform analysis further validates that the proposed adaptive gating mechanism significantly reduces unnecessary switching activity during inactive clock periods while maintaining timing integrity and functional correctness. Overall, the proposed framework demonstrates a scalable and efficient co-design methodology that integrates machine learning with hardware design, providing a robust solution for next generation low-power digital systems.
In this paper, the design and FPGA implementation of the proposed 4×4 energy-efficient Matrix Multiplication Accelerator based on Finite State Machine controlled Multiply-Accumulate architecture are presented. The implemented architecture enables efficient arithmetic operations via iterative MAC execution under sequential control operation. The presented solution provides efficient computations by means of utilizing minimal amount of FPGA hardware resources. The architecture is implemented via RTL modeling based on Verilog/System Verilog. Functional verification was accomplished using simulation to validate proper operation and arithmetic computations within all computation steps. The implementation process includes RTL synthesis, placement and routing, and timing analysis as well as FPGA-based deployment of the accelerator using Xilinx Vivado environment. Proper functionality verification revealed correct arithmetic computations and matrix multiplication operation for the implemented accelerator. After implementation and timing analysis were performed; proper timing closure was achieved with on-chip power consumption of 0.136 W and chip junction temperature of 26.6 °C. The implemented architecture utilizes only 426 LUTs, 876 flip-flops, and 16 DSP48 slices while requiring no BRAM resources. Compared to other FPGA-based accelerators, implemented matrix multiplication accelerator provides more power-efficient implementation with lower thermal dissipation and efficient computational capabilities.
Wet etching is foundational for 3D micro-/nano-structure manufacturing, yet achieving nanoscale uniformity across multiple length scales remains a persistent bottleneck. The coupled origins of etching inhomogeneity range from macroscopic gas-solid interfaces to molecular-scale dissolution reactions, and these mechanisms remain largely unclarified, hindering the development of universal precision wet processing strategies. Hydrogen silsesquioxane (HSQ), as a representative spin-on inorganic dielectric, serves as an ideal model system for investigating such cross-scale etching behaviors. Herein, we comparatively investigate HSQ wet etch-back in both acidic hydrofluoric acid (HF) and alkaline tetramethylammonium hydroxide (TMAH) systems, with TMAH-based HSQ wet etch-back demonstrated for the first time, and reveal a triple-coupling mechanism underlying HSQ etching irregularities across three distinct length scales: (i) millimeter-scale hydrogen bubbles induce hundred-micrometer to millimeter-scale under-etched protrusions via physical shielding; (ii) surface-adsorbed nanobubbles act as micro-masks and trigger micron-scale surface roughness; and (iii) a dissolution rate mismatch between cage-like monomers and self-crosslinked oligomers generates intrinsic nanoscale roughness. Guided by these mechanistic insights, we develop a targeted cross-scale regulation strategy via surfactant addition or mechanical agitation, yielding HSQ sidewall spacers with intact, gap-free sidewall interfaces on sub-30 nm diameter silicon nanopillars whose thickness is decoupled from the pillar height. The interlayer miscibility mechanism between HSQ and a photoresist is further clarified to support robust multilayer integration. Functional vertical gate-all-around MOSFETs validate process reliability. This study not only sets a universal mechanistic principle for high-accuracy wet processing of sophisticated 3D nanodevices but also provides insights into the molecular design of advanced functional photoresists targeted at 3D micro-nano manufacturing.
Recent advances in artificial deep learning-based power flow (PF) solvers have primarily focused on improving estimation accuracy, while their physical feasibility and operational reliability under realistic system variations remain insufficiently evaluated. This paper proposes a physics-informed evaluation framework for deep learning-based PF solvers under topology and load distribution shifts. Beyond conventional voltage-domain errors, the framework incorporates KCL gaps, branch-flow gaps, KVL residuals, continuous feasibility scores, and strict, practical, and relaxed PF feasibility rates. Four case studies are designed to investigate topology shifts, mixed topology training, load distribution shifts, and combined topology distribution shifts. Eight ANN architectures, including graph input models, load-only baselines, and physics-informed neural networks, are evaluated under consistent settings. The results demonstrate that low estimation errors do not necessarily correspond to physically feasible or operationally reliable PF solutions, particularly under combined system shifts. The proposed framework establishes an initial foundation and analysis for the development and verification of trustworthy deep learning-based PF solvers, providing support for their reliable deployment in future AI-enabled power system analysis.
Characterizing the performance limits of physical systems from experimental data is essential for engineering decision-making, design optimization, and reliability analysis, yet conventional regression models estimate average behavior rather than the boundary of what is achievable. This paper develops a data-driven methodology for discovering nonlinear performance boundaries from experimental observations. The problem is formulated as extreme conditional quantile estimation. A nonparametric binned percentile estimator recovers low-quantile boundary points, and the Bayesian Information Criterion selects an interpretable closed-form law from a library of physically motivated families. Non-asymptotic guarantees are stated in the fixed-design regime the experiments occupy. The method is validated on two real robotic datasets. For UR10e trajectory tracking (30,600 observations), it discovers a saturation boundary E=7.74(1−e−13.27S) mm (R2=0.962; bootstrap 95% CI for Emax of [7.05, 9.10] mm) and separates it from a steady-state envelope obeying the lag law e=tlag·v with tlag=65 ms (R2=0.997) once start-up transients are excluded. For planar pushing (40,378 pushes from the MIT MCube dataset), it discovers a power-law boundary with exponent α=2.02 (95% CI [1.88, 2.17]), indicating that the root-mean-square force fluctuation grows linearly with speed, consistent with momentum-transfer scaling at the breakdown of the quasi-static contact assumption. Family selection is stable across quantile levels and bootstrap resampling schemes, recovers the correct family in every synthetic replicate at the experimental configurations for the two family classes discovered here, and is confirmed by a prospectively specified experiment. Six of six extrapolation predictions at never-tested speeds landed within 1.2 mm, and a repeat of the protocol reproduced the per-episode 2nd percentiles to a mean absolute deviation of 0.039 mm. Compared with quantile regression, Gaussian process, support vector, and frontier-estimation baselines, the framework provides the tightest interpretable boundary estimates while approaching the target validity level, and it extrapolates where nonparametric envelopes cannot.
Inverse rendering for skin typically relies on offline path tracing to estimate physical properties, which fail to map directly to the approximate shaders utilized in interactive applications, such as game engines. We bridge this gap by directly estimating subsurface scattering parameters tailored for real-time rendering pipelines. We introduce a differentiable Screen-Space Subsurface Scattering (SSSS) framework that enables the joint optimization of spatially varying base textures and scattering distances. To account for local variations in skin thickness, we estimate these parameters across distinct facial regions, generating a blended profile map ready for production runtime engines. Our pipeline utilizes a low-cost, single-light, single-view capture setup designed for accessible and practical appearance acquisition. Optimization is driven by a masked loss function targeting regions where subsurface scattering effects are dominant. We validate our approach on a real-world dataset of $45$ individuals across the Monk Skin Tone Scale, ensuring demographic generalization. Furthermore, we complement this evaluation with synthetic data, a multi-light setup validating our single-light assumption, and a tilted-view configuration confirming angular consistency. Results demonstrate that our method effectively recovers scattering parameters from lightweight inputs, yielding realistic skin appearance directly compatible with standard real-time shaders.
A great gap in adaptivity, flexibility, and dexterity still exists between intelligent robots, whether they are made of rigid or (partially) compliant components, and their naturally evolved animal and vegetal counterparts. One reason could be of `paradigm'nature. On the one hand, most robots (both rigid and soft) are still designed by following a `Cartesian'approach, with top-down control and a clear mind-body separation. Designing artificial agents on the basis of the principles that we observe in biology is not an easy task. The principles of organization of animals and plants rely on morphological computation and emergence of fit behaviors from loosely coupled networks of embodied agents. Modeling and control require radically new approaches. In this paper we briefly recall and explain some seminal mathematical results linking information metrics and dynamics of the agents, already published years ago; we integrate them with some recent methods developed by us and by the community, we highlight some brief remarks about their relevance and we hint to our current and near future work towards finding concrete and efficient ways for the modeling and control of completely or partially compliant artificial agents and for a better understanding of physical intelligence in nature. We present a technical note summarizing phase-space volume calculations for inverted-pendulum and passive-dynamic walking models, following our proposed framework. Starting from the McGeer compass gait, we introduce knee-ankle compliance, foot-rocker effects, and a Ruina-style compass gait with distributed leg inertia. Standard heel-strike maps are derived and used in the estimation of viable phase-space volumes $W(\gamma)$, based on a state-dependent Monte-Carlo and Reproducing Kernel Hilbert Spaces methods. We show how the methods based on RKHS allow deeper and more subtle analysis.
Channel estimation is crucial in wireless communication systems. Many existing works train and test their models using the same or similar channel. However, practical channels are infinitely diverse and time-variant while data-driven algorithms often degrade on new and previously unseen channels. This prohibits the practical implementation of neural networks for channel estimation, particularly in high speed scenarios characterized by rapidly channel profile changes and very high Doppler shifts. Online training can fine-tune models to adapt to new channels, but requires high computation resource and incurs significant latency for physical-layer devices. Moreover, online training cannot exhaustively cover all possible channels because there are infinitely many possible channels. Therefore, we investigate the ability of neural networks to generalize to wireless channels. This paper proposes design criteria to generate synthetic training data, which guaranty that the resulting networks perform robustly across wireless channels, especially for very high Doppler shifts and different Doppler spectra. It ensures that offline-trained neural networks require no prior information or online training to precisely estimate the new channels. To demonstrate general applicability, we use neural networks with different levels of complexity to show that the generalization achieved appears to be independent of architecture. From simulation results, neural networks achieve a certain performance across wireless channels with different unseen power delay profiles (PDP) and Doppler shifts.
General-purpose coding agents equipped with proof-assistant tools have shown strong performance on standalone competition problems, yet cumulative textbook formalization inside a low-resource mathematical library still faces three obstacles: aligning textbook concepts with library declarations, keeping long-horizon search synchronized with live proof states, and integrating locally verified proofs reliably into a growing codebase. We present Megrez, a knowledge-guided formalization agent for Rocq libraries built on Morse--Kelley (MK) set theory. Megrez coordinates a general coding agent through three mechanisms: (1) a source-rebuildable knowledge graph merging formal declarations, compiler references, and curated metadata to support retrieval via Chinese aliases, concept clusters, proof dependencies, and multi-lane full-text ranking; (2) a feedback-driven 1--5 tactic short-burst proving loop paired with a goal-local branch ledger and exponential-interval structured reflection; and (3) a three-check acceptance pipeline (statement fidelity, axiom auditing, and full builds) coupled with a multi-corpus Provider stack and a candidate-bound recoverable batch state machine. Deployment and evaluation on a point-set topology corpus demonstrate effective cross-chapter formalization and scalable premise retrieval.
Knowledge-intensive language understanding suffers a structural reliability bottleneck: fluent answers can be produced without logically sufficient support, especially in multi-hop question answering and dynamic knowledge graphs where edge drift degrades retrieval. We introduce Counterfactual Neuro-Symbolic Proof-Carrying Memory (CNS-PCM), a backbone-agnostic neuro-symbolic harness that decouples frozen neural representations from inspectable symbolic verification checks. Candidate reasoning chains are discovered via ontology-constrained beam search (Γ) and compiled into a compact proof-carrying memory carrying crisp domain validity flags (κk), NLI-derived contradiction scores (ρk), and textual support confidences (αk). At inference time, a counterfactual verifier stress-tests prediction stability across a targeted pool of J=12 graph interventions sampled via betweenness centrality, requiring O(J·K·d) localized re-scoring operations over the K≤16 retained proofs (or O(J·d) at fixed memory capacity K, where empirical proof depth satisfies d≤3 for >94% of retained paths under a search ceiling mmax=6, bounding verifier latency overhead to ≈45 ms per query), and abstains when evidence is unstable. Rather than invasive early token fusion or prompt linearization, CNS-PCM externalizes verification into an inspectable audit record while training lightweight adapter heads using task, logical-consistency, intervention-invariance, and differentiable Brier calibration objectives. Evaluated on official public splits of HoVer, HotpotQA, and FEVEROUS with headline F1 scoring ABSTAIN as incorrect, CNS-PCM achieves test F1 scores of 80.6%, 75.2%, and 79.0%, outperforming LLaMA-2-7B + CoT (72.4%/67.1%/70.2%), GraphRAG (78.5%/73.8%/77.4%), and unthresholded Logic-LM (79.1%/74.1%/77.9%). Specifically, CNS-PCM reduces total gold-unsupported accepted assertions on the validation splits by up to 2.4 percentage points (lowering total unsupported rate from 16.4% to 14.7%, and correct-but-ungrounded assertions from 6.8% to 5.3%) and reduces Expected Calibration Error (ECE) to 6.5%, maintaining 95.2% final test coverage via a bounded re-try feedback loop (R¯=0.471 re-tries). Component ablations confirm that training invariance and inference gating provide complementary gains, while sensitivity analyses identify Stage-2 entity linking as the primary reliability lever.
Social media messages are short and noisy: the same tokens can denote a mention, a follow, or a reply depending on the surrounding user graph. Pre-trained language models frequently propose relations that a formal domain ontology forbids, whereas graph embeddings largely ignore message wording. This paper presents HyKGA (Hybrid Knowledge-Graph Attention), a single-pass neuro-symbolic encoder that (i) injects k-hop, time-decayed graph neighbors and multi-scale centrality encodings at tokenization under a strictly causal, target-edge-excised protocol, (ii) compiles the type axioms and the irreflexivity and cardinality constraints of a Description Logic ontology into two pre-Softmax masks, on self-attention and on the relation classifier, and (iii) aligns text and graph representations with an InfoNCE contrastive objective. On the Twitter Social Alignment Benchmark (TSAB), HyKGA reaches an F1 score of 0.91 and a filtered link-prediction MRR of 0.574, 5 points above QLoRA-fine-tuned LLaMA-3-8B and a late-fusion BERT–GraphSAGE hybrid, with an Ontological Violation Rate of 2.0% on relational projections (1.6% on active positive predictions and 1.0% across all thresholded predictions) against 14–35% (10.2–15.8% on positive predictions and 6.7–9.9% across all thresholded predictions) for neural baselines. Controlled ablations show that pre-Softmax masking adds 3.8 F1 points over post hoc ontology filtering. Gains of 3 to 4 macro-F1 points carry over to RumourEval, PHEME and Twitter15 as task transfer. When 20% of the axioms are removed, F1 degrades gracefully to 0.88. End-to-end latency is 18.7 ms per message, 7.6 to 11.7 times lower than retrieval-augmented pipelines on the same GPU.
Deploying deep learning segmentation models on embedded hardware requires reducing their numerical precision through quantization, an optimization whose aim is not to improve segmentation quality but to preserve it while reducing latency and energy consumption. Its effect, however, differs between architectures. This work studies whether quantization behaves the same in a convolutional network as in a Transformer, and how the compiler resolves the requested precision on two Jetson hardware–software configurations. Two ischemic-stroke segmentation models trained on the ISLES’24 dataset, M1 (convolutional nnU-Net, ∼30 M parameters, InstanceNorm) and M3 (SegFormer3D, a 4.5 M-parameter Transformer with 32 LayerNorm layers), are deployed on a Jetson Xavier NX (TensorRT 8.5) and a Jetson Orin Nano (TensorRT 10.3). Each model is compiled with TensorRT in FP32, FP16 and INT8 calibrated through the implicit (calibrator-based) workflow, with an additional mixed-precision variant for M3, and is evaluated on the same 15 test patients with the official ISLES’24 metrics (DSC, AVD, lesion-wise F1 and ALCD), latency, power, estimated energy per volume and memory, using paired per-patient statistics. FP16 preserves the quality of both models and provides most of the efficiency gain. In M1, 44% of the layers run in genuine INT8, and INT8 causes a small, statistically inconclusive DSC loss (0.169→0.154 on the Orin) but clearly more fragmented predictions (ALCD +6.9 lesions per patient), with almost no additional latency gain on the Orin. In M3, the calibrated INT8 engine keeps 97.7% of its layers in FP32, and the mixed-precision engine, with the LayerNorm layers forced to FP16, is turned by the compiler into a de facto FP16 engine (89.5% of its layers); within the tested implementation, these observations are consistent with the activation outliers that the literature associates with LayerNorm, but they do not establish it as the cause. The FP16 engine of M3 reproduces the reference (DSC 0.163) to the third decimal, and the FP32 engines of M1 match the original PyTorch model under the same pipeline, which validates the integrity of both conversions. Requesting INT8 does not guarantee INT8 execution: the effective precision is set by the compiler and depends on the model architecture. These results characterise the computational feasibility of the deployment, not the clinical readiness of the models.
Fruits and vegetables are highly perishable, and large quantities are lost after harvest through continued ripening, water loss, mechanical damage and microbial spoilage. Conventional post-harvest management depends on manual grading, visual inspection and fixed storage settings, which are subjective, slow and unable to detect internal defects or early spoilage. Artificial intelligence (AI), applied through machine learning and deep learning, has been studied as a way to make post-harvest operations faster, more consistent and non-destructive. This review examines recent advances in AI-driven post-harvest management of fruits and vegetables across four areas: quality assessment and grading; maturity, ripeness and defect or disease detection; shelf-life and spoilage prediction; and smart storage and cold chain management. Computer vision and hyperspectral imaging are used for grading and for predicting internal attributes such as soluble solids content and firmness; deep learning models detect surface and subsurface defects and classify ripeness; electronic noses combined with machine learning track spoilage and estimate remaining shelf life and the Internet of Things (IoT) combined with AI supports monitoring, prediction and traceability during storage and transport. Reported accuracies in individual studies are frequently above 90%, but several limitations restrict practical use: poor generalisation across cultivars, seasons and operating conditions; dependence on large labelled datasets; limited model interpretability; high equipment cost and the absence of shared datasets and common definitions, which makes published results difficult to compare. The review concludes that the contribution of AI to post-harvest management will depend less on further gains in laboratory accuracy and more on developing affordable, data-efficient and robust systems suited to real operating conditions, particularly for smallholders and developing countries where post-harvest losses are highest.
This study proposes a non-invasive experimental setup that integrates environmental and physiological sensors with personal parameters (age, gender, and Body Mass Index (BMI)) to measure the perceived thermal comfort of occupants via a Machine Learning (ML) framework. An experimental campaign involving 23 participants was conducted across cold (17–20 °C), neutral (22–24 °C), and warm conditions (26–28 °C). Data collected combined environmental measurements, continuous wearable physiological signals, subjective Thermal Sensation Votes (TSVs), and a multidimensional Global Sensation Vote (GSV). Evaluating six distinct datasets via a Leave-One-Subject-Out (LOSO) cross-validation strategy revealed that the value of personal data depends heavily on the underlying sensor architecture. While integrating personal data introduced slight noise in purely environmental linear models (decreasing accuracy by 4.77%), it yielded significant quantitative increments as sensor complexity shifted toward the individual, boosting accuracy by 2.03% in physiological setups and a remarkable 13.6% in bare-minimum sensor networks. Overall predictive performance reached an accuracy of 73% using a Support Vector Machine (SVM-LIN) for environmental configurations, while an XGBoost model achieved a robust 68% accuracy in a highly reduced, low-cost sensor setup by leveraging personal context. The innovative aspect of this research is the simultaneous use of sensors to monitor both physiological and environmental parameters, which, when combined with subjective feedback and personal data, significantly increases the prediction of TSV in practical cases. This integrated approach suggests prospective applications for vulnerable categories whose thermoregulatory responses deviate from standardized models, though it requires future empirical validation among these specific demographics. Overall, the integration of personal parameters into data-driven models contributes to generalizing individual physiological differences and supports the development of more effective, human-centric IoT monitoring systems.
Port Rail-Mounted Gantry Cranes are a significant investment for ports and have multiple failure modes. These cranes are complex mechanical and civil systems that run remotely without maintainers or operators present. Determining efficient methods for early fault detection is critical for optimizing that investment. We propose a solution of using commodity IoT acceleration sensors and converting the signals into the Wavelet Domain and making 3D dynamic wavelet fingerprints. This allows neural network methods, specifically autoencoder anomaly detection algorithms, to highlight anomalies within the acceleration signals. Our method is successful, utilizing a single sensor, in a noisy system of systems industrial environment, where standard signal processing and anomaly detection methods are insufficient due to the many confounding signals.
Reactive transport modelling (RTM) is essential for subsurface environmental management but is fundamentally constrained by traditional geochemical solvers. These solvers incur prohibitive computational costs and frequently suffer from numerical instabilities such as convergence failures, particularly in microbially-mediated kinetic reaction systems. While machine learning surrogates offer acceleration, they often lack physical consistency when dealing with stiff biogeochemical dynamics. Here we propose PHYGNET (Physically GNN Network), which maps microbial reaction networks into a directed graph by representing species and reactions as nodes and edges. It embeds Monod-type kinetics within a differentiable physics layer that enforces stoichiometric mass conservation and thermodynamic hierarchies, and incorporates a residual corrector that refines the final outputs at the cost of small deviations from exact mass balance. By successfully coupling with COMSOL, PHYGNET demonstrates the capability to execute full reactive transport simulations. Benchmark tests reveal that, in contrast to the severe super-linear time penalties faced by traditional solvers at engineering scales, PHYGNET maintains stable sub-linear scaling via tensor parallelism. At a scale of 10 5 grid nodes, PHYGNET achieved an acceleration (up to 3588-fold) without numerical crashes. Furthermore, its escalating speedup ratio establishes a “small-sample training, ultra-large-scale inference” paradigm that effectively offsets initial data generation costs. Overall, PHYGNET provides an efficient and physically consistent framework for accelerating Monod-type microbial reactive transport simulations, offering a practical pathway for large-scale environmental applications.
Last-mile delivery represents a critical component of global supply chains, directly influencing operational costs, customer satisfaction, and competitive advantage. As e-commerce and international trade continue to expand, companies face increasing pressure to optimize last-mile logistics to ensure timely, reliable, and cost-efficient deliveries. This examines logistics optimization models designed to enhance last-mile delivery performance while reducing operational expenses for global companies. Optimization approaches range from mathematical and algorithmic models, such as Vehicle Routing Problem (VRP) formulations and MixedInteger Linear Programming (MILP), to heuristic and metaheuristic techniques, including genetic algorithms, ant colony optimization, and hybrid methods. These models enable firms to plan efficient routes, allocate resources effectively, and adapt to dynamic demand patterns. Advancements in data-driven and AI-enabled approaches further enhance last-mile logistics by incorporating predictive analytics, IoT-enabled real-time tracking, and machine learning-based demand forecasting. Such tools allow for adaptive routing, proactive risk management, and realtime operational adjustments, thereby reducing fuel consumption, labor costs, and delivery time variability. Case studies across e-commerce, retail, manufacturing, and courier service sectors demonstrate that optimized last-mile logistics not only improve delivery speed and reliability but also enhance customer experience and brand reputation. Despite these advantages, challenges persist, including the complexity of global multi-modal networks, demand uncertainty, regulatory constraints, and technological investment requirements. Addressing these barriers requires integrated strategies combining algorithmic optimization, supplier and carrier collaboration, and digital monitoring platforms. Logistics optimization models serve as a strategic enabler for global companies, improving operational efficiency, reducing costs, and enhancing competitive positioning. By leveraging advanced optimization techniques, AI, and IoT, organizations can achieve sustainable, agile, and resilient last-mile delivery networks. This research underscores the importance of integrating technology, analytics, and strategic planning to meet the evolving demands of global supply chains, ultimately driving long-term performance and value creation.
The Internet of Things (IoT) is widely used in modern business, industry, and daily life, but its rapid advancement also brings security challenges. Processing large amounts of IoT security discussion data in a short period is difficult, especially for users who lack relevant awareness or skills. To address this issue, this study proposes an advanced MPNet-CTM-VR method using an experimental dataset collected from IoT-related literature in the Web of Science (WoS) database between 2018 and 2023. The method extracts key information for subsequent researchers to further analyze IoT and IoT security issues. The proposed model has four main components with several stages and procedures. (1) In the research topic, a masked and permuted net (MPNet) model was employed for text mining during the feature extraction stage, addressing the challenges posed by word order and polysemy. (2) For the main topic model (TM), Latent Dirichlet Allocation (LDA) was used for comparison, followed by the selection of a correlated topic model (CTM) as the core tool for text feature extraction. Afterward, we used a consistency score value (Cv) to determine the optimal number of topics for the best TM. Five well-known machine learning models, adaptive boosting (AdaBoost), convolutional neural network (CNN), decision tree (DT), random forest (RF), and extreme gradient boosting (XGBoost), were subsequently used for further classification and evaluation. (3) For the experimental validation stage, LDA and CTM processed by the bidirectional encoder representations from transformers (BERT) were compared with LDA and CTM processed by MPNet, as proposed in this study. Experimental results show that, when XGBoost is used as the classifier, the CTM processed by MPNet achieves the best evaluation results in terms of accuracy, precision, F1 score, and AUC, with 0.9508, 0.9364, 0.9361, and 0.9816, respectively. (4) Finally, VOSviewer and word clouds were used to visualize the extracted results, highlighting trends in IoT security and helping interested groups better understand the importance and challenges associated with the IoT security issue. The research findings help identify IoT and IoT security technologies, enhance knowledge of information security protection, and provide guidance for interested groups. In conclusion, this study presents a best practice with effective empirical results to benefit this related IoT and IoT security issue.
With the widespread popularization of software engineering-based Internet of Things technology, the smart home field has obtained the basic conditions to realize fine-grained monitoring and intelligent control of household energy consumption. Presently, there are two big issues with our home energy management. First, in this paper, our emphasis is only on conserving energy rather than considering whether the energy we consume from the grid is clean or dirty. Second, we are not making sure that the software we use to schedule our energy use is reliable. To address these issues, this study proposes a carbon-aware home energy management framework that integrates predictive modeling, environmentally conscious scheduling, and software reliability validation. The three major entities of the proposed scheme are predicting the closeness, predicting energy usage, and scheduling to utilize energy to reduce our carbon footprint. In this context, utilizing a special formula that considers both energy consumption and carbon emissions, we can find ways to make our homes better and more efficient for the planet. For this newly proposed system, all simulation experiments were completed on the OMNeT++ platform. After the simulation tests were finished, several types of reliability verification tests were also conducted. The results of these tests are clear: all core performance indicators of this new framework outperform those of the benchmark method used as a reference. This system can provide an accurate, long-term usable solution for next-generation smart home energy management systems. Received: 6 May 2026 | Revised: 1 July 2026 | Accepted: 9 September 2026 Conflicts of Interest The authors declare that they have no conflicts of interest to this work. Data Availability Statement The data that support the findings of this study are openly available in the UCI Machine Learning Repository at https://archive.ics.uci.edu/ml/datasets/individual+household+elect ric+power+consumption. The grid carbon-intensity data are openly available from the National Grid Carbon Intensity API at https://carbonintensity.org.uk/. The OMNeT++ simulation configurations, preprocessing scripts, and source code used in this study are available from the corresponding author upon reasonable request. Author Contribution Statement Ali Raad Abdulkareem: Conceptualization, Methodology, Investigation, Writing – original draft, Project administration. Mustafa Moosa Qasim: Conceptualization, Methodology, Software, Formal analysis, Data curation, Writing – original draft, Visualization. Husham Lateef Swadi Roomi: Software, Investigation, Data curation, Visualization. Murtadha Al-Maliki: Investigation, Resources. Mahmood A. Al-Shareeda: Conceptualization, Methodology, Validation, Formal analysis, Writing – review & editing, Supervision, Project administration. Mohammed Amin Almaiah: Validation, Formal analysis, Resources, Writing – review & editing, Supervision, Funding acquisition. Rami Shehab: Validation, Resources, Writing – review & editing, Funding acquisition.
Artificial intelligence (AI) is increasingly integrated with smart sensors, Internet of Things (IoT) networks, machine learning, deep learning, and computer vision to support site-specific agricultural management. However, strong performance on computational benchmarks does not necessarily translate into reliable decisions, effective field interventions, or measurable improvements in agricultural productivity and resource efficiency. This comprehensive narrative survey critically synthesizes advances in agricultural sensing, IoT architectures, data processing, multimodal sensor fusion, predictive modeling, computer vision, decision optimization, and automated actuation. Particular emphasis is placed on the relationships between measurement reliability, model generalization, decision feasibility, intervention accuracy, and verified agricultural outcomes. The reviewed literature demonstrates promising capabilities in crop monitoring, yield prediction, disease and weed recognition, precision irrigation, and selected automated operations. Nevertheless, heterogeneous evaluation protocols, limited independent field validation, sensor calibration challenges, distribution shifts, interoperability barriers, and deployment costs constrain operational reliability and generalizability. To address these interconnected challenges, the survey proposes a literature-informed reference architecture linking agricultural observations, data integration, AI inference, constrained decision-making, physical interventions, and outcome feedback, supported by uncertainty assessment, human oversight, and cross-layer assurance. The proposed architecture is conceptual and has not been empirically validated. Future research priorities include reproducible field evaluations, representative agricultural datasets, uncertainty-aware and explainable AI, resource-efficient computing, interoperable infrastructures, and economically inclusive deployment. The review concludes that the practical value of AI-enabled precision agriculture should be evaluated through reliable interventions, sustainable resource utilization, and demonstrated agricultural outcomes rather than predictive performance alone.
Bluetooth Low Energy (BLE) devices are ubiquitous across consumer, industrial, and healthcare IoT ecosystems, yet their resource-constrained nature typically precludes conventional cryptographic authentication, leaving device identification as a security primitive for device management, access control, and threat detection in IoT environments. Current BLE fingerprinting techniques are restricted by rigid cryptographic hashes, machine learning models that require periodic retraining to cover new devices, or single-layer feature sets vulnerable to missing data. In response to these limitations, we introduce a Nilsimsa-based identification approach that performs BLE device type/manufacturer identification through merging link-layer features and advertised GATT service information before network access is granted to cope with missing or inaccessible data. We analyzed various combinations of features from different layers to assess how merging them impacts identification accuracy. The exhaustive analysis uses 10 test objects and 10 reference hashes. The large-scale evaluation uses 5000 test records and 10,000 reference hashes covering 104 device type–manufacturer pairs. Key findings highlight that cross-layer features are important to cope with missing features. The minimum subset size was found at k = 10 with a subset-coverage criterion: at least 95% of all subsets of 10 features reach an accuracy of 90% or more. Our proposal not only facilitates the seamless enrollment of new devices through reference-hash registration without retraining but also enables confidence-based anomaly detection using its similarity score, providing an incrementally extensible approach for BLE security monitoring. Overall, our approach advances BLE fingerprinting by introducing a schema-agnostic, retraining-free, cross-layer identification model tailored for the dynamic environment of IoT deployments.
Federated Learning (FL) is emerging as a distributed privacy-preserving paradigm, aiming to build a strong machine learning (ML) model in a distributed manner. Client selection in FL is a critical challenge for building intrusion detection systems, especially in heterogeneous Internet of Things (IoT) networks, where Non-Independent and Identically Distributed (Non-IID) data partitioning causes classical selection strategies to suffer from catastrophic forgetting of minority attack classes. This paper presents a Quantum Approximate Optimization Algorithm (QAOA) based framework that formulates client selection as a Quadratic Unconstrained Binary Optimization (QUBO) problem encoding client utility, communication latency, a cardinality constraint, and an optional group diversity term, solved at each communication round using the QAOA. To reduce quantum circuit width for near-term hardware compatibility, we investigate two qubit reduction strategies in this paper: hierarchical client clustering and classical pre-filtering. Extensive experiments on three IoT intrusion detection benchmarks under heterogeneous Non-IID distributions show that our proposed approach improves macro F1-score by up to 30% over Greedy-TopK selection and by up to 29% over random client selection. Our ablation studies confirm that the quantum-assisted selection advantage emerges specifically under Non-IID distributions. On the most challenging dataset, where two attack categories represent less than 0.5% of the data, the proposed method achieves superior results than other client selection methods on aggregate metrics while offering superior precision and convergence stability.
Biological intelligence emerges from neural activity that connects sensation, memory and action. Synaptic plasticity, neuronal assemblies and oscillations provide principles for biomimetic computing, but translating them into reusable operations requires a defined computational unit. We introduce Oscillatory Processing Units (OPUs), groups of synthetic neurons sharing a rhythm. Their directed interactions transform incoming activity and prior receiving state into an Oscillatory Occupancy Space (OOS), which becomes available to subsequent computation. We show that source--receiver conditions and oscillatory organization shape these transformations, including their geometry, persistence and historical accessibility. An Atlas of measured relationships supports prospective composition with predicted response classes. Configured circuits perform logical and numerical operations, express associative recall and generate closed-loop decisions with fixed synaptic weights. OPUs and OOS establish a computing substrate for biomimetic intelligence in which experience remains part of the state through which new input becomes action.
Visible and near-ultraviolet light underpins applications from biophotonics to quantum computing, which demand fast control of optical amplitude and phase. Integrated photonics offers scalable control, yet most visible-wavelength platforms lack gigahertz-scale modulation. We demonstrate a thin-film lithium niobate platform at 493 nm integrating grating couplers, beamsplitters, ring resonators, and modulators. Segmented traveling-wave modulators achieve a half-wave voltage--length product of 450 mV$\cdot$cm, $>$20 dB extinction, and bandwidth beyond 50 GHz. Ring resonators show $2.9 \pm 0.2$ dB/cm propagation loss and quantify the platform's photorefractive response. This enables agile visible beam delivery for trapped-ion systems, although photorefraction and substrate uniformity remain key challenges to scaling.
Dynamical codes protect quantum information with sequences of measurements, generalizing the stabilizer formalism of static codes. However, the dynamical formalism has been developed exclusively for qubit systems. While Gottesman-Kitaev-Preskill (GKP) codes are promising candidates for quantum computing that encode qubits in the continuous phase space of oscillators, no dynamical formalism has been developed for them. In this work, we bridge this gap and introduce dynamical GKP codes, combining continuous-variable lattice encodings with time-dependent translation measurement dynamics. To obtain the update rules and the distance of a given measurement sequence, we first generalize GKP codes to subsystem GKP codes using scaled symplectic lattices. Building on this structure, we adapt the ancestry, detector, and benign-error machinery originally developed for dynamical qubit codes. From this construction, we obtain a general and efficient method to compute the syndrome and distance of a dynamical GKP code from a persistent stabilizer lattice and a bounded inference window. We illustrate the framework with the example of the two-mode square-hexagonal Floquet GKP code. This code has reduced measurement weight compared to both the square and hexagonal GKP codes, while combining their set of logical gates, resulting in easier implementation of the full single-qubit Clifford group. Finally, we show that measurements with irrational symplectic overlap, which have no analog in discrete qubit codes, reduce to destructive homodyne detection in the single-mode case. These results extend an important qubit-code tool to continuous-variable systems and identify a new bridge between discrete stabilizer measurements and continuous homodyne detection.
Objective: Epidemic outbreaks are often asynchronous across regions, yet collaborative forecasting is hampered by privacy laws and heterogeneous public health interventions. We developed and evaluated, as a proof of concept, a federated learning (FL) framework for cross-regional COVID-19 forecasting under asynchronous outbreaks among Guangdong, Macau, and Portugal, while maintaining data sovereignty. Methods: Using daily case data from WHO and China CDC for the three regions, whose outbreaks were asynchronous and interventions heterogeneous, we trained Long Short-Term Memory networks via FL for short-term forecasting, benchmarked against localized and centralized training using Root Mean Square Error, Mean Absolute Error, and Symmetric Mean Absolute Percentage Error. All experiments used the Flower FL framework with consistent hyperparameters across sites. Publicly available, aggregated case-count data were used so that the centralized benchmark could be computed as a reference. Results: The dataset comprised 1,233 daily observations (411 days) of new confirmed COVID-19 cases from November 24, 2021, to January 8, 2023. The earlier Omicron wave in Portugal supplied transmission dynamics that improved forecasts in the later-onset regions of Macau and Guangdong. The FL model achieved a normalized Root Mean Square Error of 0.028, an 88.3% improvement over the local model (0.240) and comparable to the centralized model, with network overhead mainly determined by model size and aggregation rounds. Conclusion: FL achieved short-term forecasting accuracy close to centralized training while maintaining data residency, with acceptable overhead. Leveraging asynchronous outbreaks across regions may offer a practical path to privacy-preserving collaborative forecasting, though prospective validation is required.
Anomaly-based malware detectors learn what benign files look like and flag deviations. Generative approaches to this problem can make use of a"latent"or"prior"data representation, which is often chosen to be a Gaussian distribution. We ask whether the choice of prior distribution matters, and whether one produced by quantum hardware behaves differently from a classical one. Building on SAGEGAN, a style-based detector trained only on benign data, we compare three latent distributions: the original 512-dimensional Gaussian, a 48-dimensional Gaussian, and a 48-dimensional distribution sampled from an ORCA Computing PT-2 photonic quantum processor. We then refine each trained model's frozen representation using greedy and reinforcement-learning methods. Across ten random seeds, the quantum-derived model shows richer internal structure with better effective rank and latent similarity. It also benefits most from refinement. Though final detection accuracy is similar to that of the refined Gaussian latent, these findings show differences in latent geometry and refinement behavior between the trained configurations.
We consider multi-stage stochastic optimization problems with smooth convex stage costs and linear linking constraints. Dynamic stochastic approximation (DSA), a stochastic first-order method, uses a recursive structure whose memory requirement grows only linearly with the number of stages. Its implementation, however, requires an exact solution of the proximal subproblem associated with the stage cost, and this subproblem often admits no closed-form solution. To resolve this issue, we develop an accelerated dynamic stochastic primal-dual (A-DSPD) framework that retains the recursive value-function structure and only requires the gradient of the stage cost. Within this framework, each stagewise saddle-point subproblem is solved by an inexact accelerated stochastic primal-dual method (iASPD), which applies Nesterov acceleration to the smooth stage cost and uses recursively generated stochastic $\epsilon$-subgradients for the nonsmooth downstream value function. For the stochastic composite saddle-point problem, after $N$ primal-dual updates, iASPD yields optimality and feasibility bounds in which the contribution of the smooth stage cost decays at a faster rate $O(1/N^2)$, while the terms associated with stochastic $\epsilon$-subgradients of the nonsmooth value function decay as $O(1/\sqrt{N})$, up to the prescribed oracle inexactness. We also develop bounds for the stochastic $\epsilon$-subgradient estimators and the dual iterates, which yield complexity guarantees for A-DSPD. Specifically, A-DSPD computes an $\epsilon$-solution using $O(1/\epsilon^4)$ sampled scenarios and $O(1/\epsilon^5)$ primal-dual updates for three-stage problems. We further extend the method and its convergence analysis to problems with an arbitrary finite number of stages. Numerical experiments on three-stage stochastic portfolio models illustrate the empirical advantage of the proposed method.
Digital quantum simulation of open quantum systems has produced several methods for reproducing non-Markovian dynamics, among them pseudomode embeddings, the hierarchical equations of motion, and collision models. These methods are validated against populations, coherences or spectra, not against the measures that quantify memory. We study how the Breuer-Laine-Piilo (BLP) measure behaves under the approximations of a digital simulation. The circuits run on classical density-matrix simulators, where the exact value of the measure is computable. First, a two-level encoding of the pseudomode misrepresents the measure with a regime-dependent sign. It overstates the memory by a factor of 2.3 at strong coupling, loses most of it at moderate temperature, and both signs occupy large regions of the coupling-temperature plane. A variance-aware rule for the required Fock dimension, ⌈⟨n⟩ + cσₙ⌉ + 1, accounts for the number statistics: at equal mean occupation, Fock, coherent and thermal states need 4, 8 and 16 levels. We calibrate c for trajectory- and measure-level targets and cross-validate it across three families. Second, the raw finite-shot estimate overstates the true value by up to 89% under realistic noise. A significance-threshold estimator reduces this bias to −0.004 at the strongest noise, and a Markovian control with a percentile rule keeps false positives at or below 0.3%, with measured miss rates. Third, slow non-Gaussian device noise creates backflow in a memoryless channel, so the measure characterizes the implemented channel and not the source of its memory. The Rivas-Huelga-Plenio measure, computed from reconstructed maps, does not converge on these channels and serves only as a diagnostic. At equal two-qubit gate count, a pseudomode and a two-site collision chain cross over in accuracy. Every result is validated against an exact reference and reproducible from recorded data. We make no claim of quantum advantage.