This paper proposes a novel approach to computing, termed Dynamic Semantic Network Neuro-Morphic Computing, which leverages the principles of biological neural networks to achieve parallel, adaptive learning, and reasoning for complex data structures. The core idea is to mimic the dynamic connectivity and synaptic plasticity mechanisms found in biological neurons, creating a programmable neuro-morphic architecture. This architecture utilizes dynamically adjustable neural network connections and synaptic strengths based on input data and learning algorithms. A temporal signal processing and feedback mechanism simulates the dynamic behavior of biological neural networks, combined with Graph Neural Networks (GNNs) to construct and update dynamic semantic networks for representing and inferring relationships within data. This approach addresses the limitations of existing neuro-morphic computing, which primarily focuses on static hardware architectures, and traditional GNNs facing inefficiencies in handling large-scale, dynamic semantic networks. The resulting system aims to provide real-time learning and reasoning capabilities for complex data relationships, with potential benefits of low power consumption and high parallelism. The system is formally defined as follows: Let *S* = {*s*1, *s*2, ..., *s*K} be a set of nodes representing data elements. Let *E* = {*e*1, *e*2, ..., *e*N} be a set of edges representing relationships between nodes. Let *W* = {*w*ij} be a matrix representing the connection weights between nodes *s*i and *s*j, where *w*ij ∈ ℝ. Let *θ* = {*θ*ij} be a matrix representing the synaptic strengths between nodes *s*i and *s*j, where *θ*ij ∈ ℝ. Let *a*i ∈ ℝd be the activation value of node *s*i at time *t*. Let *l*i ∈ ℝ be the learning rate for node *s*i at time *t*. Let *h*i ∈ ℝd be the hidden state of node *s*i at time *t*. The dynamic update rule for node activation is given by: *a*i(t+1) = σ(*∑*j (*w*ij*h*j(t+1)) + *θ*ij *a*i(t+1)) where σ is an activation function (e.g., sigmoid, ReLU). The dynamic update rule for hidden state is given by: *h*i(t+1) = *f*(*a*i(t+1)) where *f* is a function that transforms the activation value into a hidden state. The learning rule updates the connection weights and synaptic strengths as follows: *w*ij(t+1) = *w*ij(t) + *l*i *∑*k (*w*ik(t+1) (*a*k(t+1)) ) *θ*ij(t+1) = *θ*ij(t) + *l*i *∑*k (*w*ik(t+1) (*a*k(t+1)) ) where *l*i is the learning rate. The system's performance is evaluated based on metrics such as accuracy, convergence time, and energy consumption. The core architecture will be implemented using a neuromorphic hardware platform, potentially utilizing spiking neural networks (SNNs) for efficient temporal processing.
Jincheng Zhang· Zenodo (CERN European Organi...· 0 citations
The computing continuum extends computation beyond centralized cloud data centers to edge resources, supporting applications with stringent Quality-of-Service (QoS) requirements. However, the heterogeneity and dynamic nature of edge environments make orchestration challenging, particularly when decisions rely on static network information. This paper proposes a network-aware orchestration architecture that integrates a continuously maintained network representation to support adaptive application placement and management. By combining network and computational awareness, the proposed approach aims to improve orchestration decisions and application performance across the computing continuum.
Fernando Rego, José Eduardo Ferreira da Costa, Carlos Resende et al.· Zenodo (CERN European Organi...· 0 citations
ABSTRACT Multi‐valued logic (MVL) offers a promising approach to overcoming the intrinsic limitations of binary computing by enhancing data density at the device level. However, achieving stable and well‐defined multi‐level logic states within a single semiconductor channel layer remains a significant challenge. In this work, we demonstrate a ternary inverter based on a two‐step doping mechanism in a single polymer semiconductor channel incorporated into an organic electrochemical transistor (OECT). The device exhibits a robust intermediate conduction state characterized by zero‐differential transconductance (ZDT). The face‐on molecular orientation and dense alkyl side‐chain packing initially impede ion penetration, giving rise to surface‐confined electrostatic gating at low gate voltages ( V G ), before transitioning to bulk volumetric electrochemical doping at higher V G . This sequential doping mechanism yields a wide and stable ZDT plateau, an effect that is less pronounced in polymer semiconductors adopting edge‐on orientations or amorphous microstructures. This work highlights the critical role of molecular orientation and side‐chain packing in realizing distinct multi‐level logic states, offering design guidelines for MVL circuits.
Dong Hyun Park, Min Su Kim, Jae Hyun Han et al.· Advanced Functional Material...· 0 citations
This paper presents a novel algorithm for dynamic topology optimization, leveraging the unique properties of quantum computing to dynamically adjust network parameters. The core mechanism utilizes quantum entanglement and superposition to efficiently compute and optimize the topology of a network, leading to improved performance compared to traditional optimization techniques. We propose a quantum-enhanced approach to network design, focusing on the dynamic adjustment of node and edge weights to achieve optimal network behavior. This work explores the potential of quantum computation to address challenges in network routing, congestion control, and resource allocation, offering a promising avenue for enhancing network efficiency and resilience. The proposed algorithm demonstrates improved results through extensive simulations and theoretical analysis, highlighting the potential for significant advancements in network management.
Jincheng Zhang· Zenodo (CERN European Organi...· 0 citations
Reach audiences
Advertise in front of researchers, engineers, and readers.
Spin-torque diodes based on magnetic tunnel junctions have emerged as a promising paradigm for microwave detection and neuromorphic computing due to their intrinsic nonlinear electrical characteristics. Traditional artificial intelligence image processing technology based on the von Neumann architecture faces the bottlenecks of power consumption and storage wall. Therefore, we propose and demonstrate hardware-inspired artificial neurons featuring a rectified linear unit activation function by leveraging injection-locked spin-torque diodes. Taking full advantage of the above properties, we construct a supervised autoencoder architecture for image denoising. The proposed architecture is validated on the MNIST dataset with additive Gaussian noise. Experimental results show that its denoising performance is on par with traditional software-implemented neural networks, featuring similar convergence performance and stable high-quality image reconstruction. Notably, even under heavy noise intensity of 0.3, the system consistently maintains a peak signal-to-noise ratio above 20 dB and a structural similarity index exceeding 0.85 across all scenarios. This study provides a viable hardware-native technical pathway toward edge artificial intelligence applications.
Yazhong Si, Fuqian Ge, Like Zhang et al.· Applied Physics Letters· 0 citations
This paper proposes a novel approach to computing leveraging dynamic topology neural-spike networks. The core idea centers on mimicking the self-adaptive topology structures found in biological neural networks to achieve higher efficiency and robustness in complex computational tasks. We introduce a programmable hardware platform composed of simulated neurons with dynamic connections and synaptic plasticity. Utilizing machine learning algorithms, specifically reinforcement learning, we continuously optimize the network topology in response to task demands, encompassing node addition, removal, and weight adjustments. This dynamic adaptation allows the network to real-time adjust to fluctuating input data, realizing adaptive computation. The innovation lies in the *dynamic* topology, contrasting with static structures or simplified models in existing neural-spike computing systems. By integrating parallel processing with machine learning optimization, our framework promises enhanced computational efficiency and resilience, representing a significant advancement over conventional neural-spike computing paradigms. The key mathematical framework revolves around representing the network topology as a graph (G = (V, E)), where V is the set of nodes (neurons) and E is the set of edges (connections) with associated weights. The dynamics of the network are governed by the following stochastic differential equations: d*s*i/dt = ∑j∈N(i) *w*ij *s*j + *f*i, where *s*i is the state of neuron *i*, *w*ij is the synaptic weight connecting neuron *i* to neuron *j*, *N(i)* is the set of neurons connected to neuron *i*, and *f*i represents a stochastic input or intrinsic noise. The learning process is formulated as a Markov Decision Process (MDP), and the policy is learned using reinforcement learning algorithms, aiming to maximize the expected reward. The core of the system can be represented as: R = ∑i αi *s*i, where αi is the activation function of neuron *i*. The system is designed to minimize the error between the output and the desired output, using a cost function: E = ∑i || *s*i - *t*i||2.
Jincheng Zhang· Zenodo (CERN European Organi...· 0 citations
CeO₂ is key in catalysis, fuel cells, chemical mechanical planarization, and oxygen storage due to its Ce⁴⁺ / Ce³⁺ redox couple and oxygen mobility. At the nanoscale, properties become morphology-dependent, yet the link between morphology and mechanical behavior‒particularly elastic properties and defect structure‒remains underexplored. Here we investigate four CeO₂ morphologies (nanoparticles, spherulites, intermediate rods, and long rods) using XRD line profile analysis, bond valence sum (BVS) calculations, and DFT+U. Rietveld-refined lattice parameters serve as direct DFT inputs. Four Williamson-Hall models quantify crystallite size, microstrain, dislocation density, and character, with DFT-computed direction-dependent Young’s moduli and contrast factors. Results show bulk (≈169 GPa), shear (≈58.8 GPa), and Young’s (≈158 GPa) moduli remain constant within uncertainty across morphologies. The high dislocation densities (up to 10¹⁵ m⁻²) represent a small volume fraction (≈0.03 %) and do not appreciably alter the effective polycrystalline moduli. In contrast, defect structure varies dramatically: dislocation density increases from 10¹¹ to 10¹⁵ m⁻², transitioning from screw-dominated to mixed / edge character with aspect ratio. Oxygen displacement parameters (Uiso, O = 0.0136 – 0.0669 A²) peak in rods, indicating enhanced anionic disorder, while BVS (4.71– 4.75 v.u.) reveals compressed Ce-O bonds. These findings demonstrate that defect populations are morphology-dependent, enabling independent optimization for catalysis, chemical mechanical planarization (CMP), solid oxide fuel cells (SOFCs), and coatings, while the mechanical backbone of CeO2 remains unchanged.
Balter Trujillo-Navarrete, F. Paraguay‐Delgado, Rosa María Félix-Navarro et al.· National repository of resea...· 0 citations
This paper investigates the application of neuromorphic computing to address the limitations of traditional von Neumann architectures in real-time image processing. The core argument presented is that the inherent bottleneck of separating memory and processing units in conventional computers significantly hinders the performance of computationally intensive image processing tasks. We propose leveraging neuromorphic chips and spiking neural networks (SNNs) as a solution. SNNs, inspired by the biological brain, offer the potential for drastically reduced power consumption and increased processing speeds compared to conventional digital systems. This work outlines the theoretical framework and key considerations for implementing image processing algorithms directly on neuromorphic hardware. Specifically, we explore the mapping of common image processing operations, such as edge detection and feature extraction, onto SNNs. The potential for achieving near-real-time performance with minimal energy expenditure is highlighted. Further research directions, including optimization strategies for SNN design and exploration of different neuromorphic architectures, are discussed.
Jincheng Zhang· Zenodo (CERN European Organi...· 0 citations
pyunwrap unwraps InSAR phase using a physics-informed U-Net that predicts the integer ambiguity map rather than the unwrapped phase itself. The core identity unwrapped = wrapped + 2π·k : is enforced by construction, so the model is structurally incapable of contradicting the observed wrapped phase. This is the first tagged release: the full pipeline from synthetic data through training, tiled inference, analytics, and reporting is implemented, tested, and documented end to end. Highlights AmbiguityNet : a ResNet-34 U-Net predicting integer phase ambiguity via a straight-through-estimator rounding head, with an auxiliary residue-probability head for uncertainty. Realistic synthetic data : Gaussian bowls, a from-scratch Okada (1985) fault dislocation model, a Mogi (1958) volcanic source, DEM topography, Kolmogorov atmospheric noise, orbital ramps, decorrelation noise, and pseudo-real ALOS-2 rewrapping. Curriculum training with optional SNAPHU pseudo-ground-truth fine-tuning on real data. Production tiled inference with edge-aware, residue-weighted smart merging of the ambiguity map , never the phase directly , plus Monte Carlo Dropout uncertainty and an ONNX Runtime → OpenVINO fallback chain. Full analytics suite : residue detection, Nyquist gradient analysis, Grad-CAM, Integrated Gradients, uncertainty calibration, and an automated self-contained HTML report. 65 tests, including known-answer physics tests and a tile-merging regression test that catches boundary-artifact bugs numerically, not visually. Two pre-executed example notebooks covering the training chain and the full end-to-end pipeline against real synthetic data. What's included Synthetic data engine Four deformation models (Gaussian bowl, Okada, Mogi, deformation-free control), DEM-driven topographic phase, Kolmogorov-spectrum atmospheric turbulence, orbital ramps, and coherence-dependent decorrelation noise. Ground-truth ambiguity is computed against the actual observed (noisy) phase ,unwrapping resolves 2π ambiguity, it does not denoise ,which keeps every generated sample exactly consistent with wrapped + 2π·k == unwrapped to floating-point precision. Data pipeline Full-coverage sliding-window tiling, percentile-robust amplitude normalization, and an InSARTileDataset with 8-fold dihedral augmentation that recomputes the integer ambiguity map after every transform rather than reusing a cached value, so it stays exact even at the ±π wrap boundary. Model & training AmbiguityNet (24M params) plus PhysicsInformedUnwrapLoss, a four-component loss combining supervised ambiguity regression, re-wrap consistency, coherence-weighted smoothness, and an ambiguity-map residue penalty. Trainer implements three-stage curriculum learning, AdamW with warmup/cosine annealing, gradient clipping, TensorBoard logging, and optional SNAPHU-based fine-tuning. Inference & deployment PhaseUnwrapper tiles arbitrarily large interferograms and merges results with an edge-aware taper (only feathering edges that actually border a neighboring tile , a plain Hanning window incorrectly zeroes scene boundaries too) weighted by the model's own residue-probability output. ONNX export is single-file by design, with GPU → CPU → OpenVINO backend fallback. Analytics, visualization & reporting Goldstein-style residue detection and clustering, Nyquist gradient- violation mapping, error-distribution statistics, Grad-CAM and Integrated Gradients explainability, uncertainty-calibration reliability diagrams, interactive Plotly 3D surfaces, a folium swipe-comparison map, and a Jinja2-templated HTML report ,all wrapped in best-effort error handling so a reporting failure never breaks a production inference call. Testing & CI A 65-test suite (pytest) covering synthetic generation, model/loss correctness, physics/analytics correctness against hand-constructed known-answer fields (e.g. an exact phase vortex with a known topological charge), and full-pipeline integration. GitHub Actions runs a fast-test matrix across Python 3.10–3.12, a separate slow/integration job, and a ruff/black lint job. Installation git clone https://github.com/EOCoreINT/pyunwrap.git cd pyunwrap pip install -e ".[dev,maps,deploy,notebooks]" See the README for the quickstart and the architecture reference for how data flows through the pipeline. Known limitations This is an early-stage release and is honest about where it currently falls short: No pretrained weights ship with this release. PhaseUnwrapper.from_pretrained() downloads from a Zenodo record you supply; there is no benchmarked checkpoint yet. Monte Carlo Dropout uncertainty is currently a no-op. AmbiguityNet has no nn.Dropout layers yet (only BatchNorm), so uncertainty in UnwrapResult is identically zero. The residue-probability head is the meaningful uncertainty signal available today. No published accuracy benchmark against SNAPHU or other classical unwrappers yet. The physics-consistency guarantees are proven (exhaustively, in tests); real-world unwrapping accuracy on held-out Sentinel-1 data has not yet been formally evaluated in this repo. APIs may change between minor versions until 1.0. Acknowledgments Built on the shoulders of the classical InSAR literature this package's synthetic models and design decisions are grounded in: Goldstein, Zebker & Werner (1988); Itoh (1982); Chen & Zebker (2001, SNAPHU); Okada (1985); Mogi (1958). Full citations in the README. Full Changelog See CHANGELOG.md for the complete, itemized history, including every bug found and fixed during development.
Samuel Appiah Kubi· Zenodo (CERN European Organi...· 0 citations
This paper investigates the potential of neuromorphic computing for event-driven data processing. Traditional computing architectures often struggle with the inherent inefficiencies of handling continuous data streams, leading to significant energy consumption. The core claim presented here is that mimicking the brain's event-driven neural processing offers a pathway to dramatically improved data processing efficiency and reduced energy expenditure. The proposed approach centers around constructing a neuromorphic system utilizing temporal circuits, leveraging the principles of event-driven signal transmission and parallel computation to achieve real-time data stream processing. This work outlines the theoretical framework and key design considerations for such a system, highlighting its advantages over conventional approaches. The research addresses the critical need for more energy-efficient data processing solutions, particularly in applications such as sensor networks, edge computing, and real-time analytics. The system's ability to react only to significant events within the data stream, rather than processing every element, is expected to yield substantial performance gains. Ultimately, this exploration contributes to the growing field of neuromorphic computing and its potential to revolutionize data handling paradigms. ---
Jincheng Zhang· Zenodo (CERN European Organi...· 0 citations
This article proposes a systematic improvement path for the security governance capability of distribution networks driven by intelligent technology. Realize global state awareness through IoT sensors and edge computing; Based on AI algorithms for fault diagnosis and situation deduction, combined with digital twin platform optimization decision-making; Constructing self-healing control system based on intelligent terminal and virtual power plant technology. The technical level focuses on cloud-side collaborative architecture and deep learning algorithm, the management level promotes cross-departmental collaboration and process standardization, and the mechanism level improves the standard specification and dynamic evaluation mechanism. Taking a demonstration area as an example, after the implementation, the fault location time is reduced from 42 minutes to 2 minutes, the annual power outage events are reduced by 78%, and the voltage qualification rate is increased to 99.71%, which verifies the effectiveness of the closed-loop system of "perception-analysis-decision-execution". The research shows that the integration of intelligent technology, management innovation and institutional guarantee can improve the operation toughness and governance efficiency of distribution network.
What if pathology foundation models could do more with less? GigaPath-Flash and GigaTIME-Flash cut computational demands while maintaining strong performance, opening the door to larger studies and broader exploration. The post GigaPath-Flash and GigaTIME-Flash: Toward population-scale discovery with efficient pathology foundation models appeared first on Microsoft Research.
MIT News · Artificial Intelligence· news.mit.eduAug 31, 2026
With millions of users across the world, Julia has been used to conduct cutting-edge research and to design new drugs, jet engines, heat pumps, and more.
MIT News · Artificial Intelligence· news.mit.eduAug 27, 2026
A new machine-learning framework aims to improve the success rate of computational protein design while moving away from results that reproduce sequences found in nature.
