Total impulse characterizes the overall propulsive capability of a solid rocket motor. The cost and resources required for ground static firing tests limit data availability, while temperature-dependent thrust variations complicate modeling with few records. This study develops an MT model that combines model-agnostic meta-learning with a Transformer. Complete thrust records are divided into three segments using the thrust peak and an amplitude threshold, then downsampled. The model receives the compressed sequences and their actual timestamps, extracts temporal relationships through self-attention, and learns a shared initialization from high- and low-temperature tasks. This initialization supports total impulse prediction from a small set of ambient-temperature records. Model comparisons, sampling analyses, and cross-temperature validation assess the use of compressed thrust information and historical operating-condition data. With three-segment, 500-point inputs, MT achieves a mean absolute percentage error (MAPE) of 0.429511% on the ambient-temperature query set. Record-level error analysis and cross-temperature validation show that information learned from high- and low-temperature tasks improves ambient-temperature total impulse estimation with limited target records. Combining sequence representation with task adaptation improves the use of existing temperature-specific test data for few-shot total impulse prediction and motor performance analysis from compressed records.
Personalized pronunciation-exercise design requires context-sensitive content, longitudinal adaptation, and professional oversight. This study presents Speak4All, a cloud-deployed human-in-the-loop (HITL) generative AI framework integrating structured learner profiles, professional assessments, interpretable difficulty recommendations, deterministic constraint validation, and explicit human control. Following a Design Science Research approach, an ex ante evaluation examined scenarios involving children with Down syndrome without recruiting children as participants. Two complementary settings were considered: an exploratory assessment with eight adult stakeholders (two speech-language professionals and six caregivers) in Ecuador and a remote controlled evaluation involving 25 education professionals and 25 structured cases. Compared with prompt-only and profile-aware generation, the integrated framework achieved higher educator ratings and 84.8% acceptance without editing across exercise-evaluation decisions. All retained exercises satisfied the specified hard constraints (100.0% post-validation compliance). Adaptive recommendations achieved 87.0% exact agreement with independent educator judgments (κw=0.848), with 85.0% accepted without modification. Stakeholder feedback indicated favorable perceptions of personalization, progress monitoring, and multimedia support. These findings support the framework’s ex ante technical feasibility and potential usefulness for professional-supervised educational decision support, without establishing clinical effectiveness. Future research should evaluate supervised use with children through validated speech-related outcome measures.
Cross-subject motor-imagery electroencephalography (MI-EEG) classification remains challenging due to substantial inter-subject variability and non-stationarity, which cause task-relevant spectral and spatial patterns to vary across users and reduce the robustness of models trained on source subjects. To address this issue, we propose NeuroMambaNet, a brain-informed dual-stream network that explicitly integrates spectral, spatial, and rhythmic priors. In the frequency-enhanced stream, a Mamba-Augmented Multi-Band Oscillation (MAMBO) module decomposes EEG with a Discrete Wavelet Transform (DWT), models the physiologically relevant θ, α, β, and γ sub-bands using independent Mamba blocks, and reconstructs a band-enhanced signal through inverse DWT (IDWT). In parallel, the raw stream extracts complementary spatial representations using two Brain-Region Prior Grouping (BRDW) branches based on left–central–right and central–outer electrode partitions. β-band evidence from frontal-central electrodes is then used by a lightweight Beta-Gate module to adaptively weight the two spatial branches. The frequency-enhanced and gated raw representations are fused at a later stage and processed by a global temporal convolution for classification. Under a strict leave-one-subject-out (LOSO) protocol, NeuroMambaNet achieves an average accuracy of 59.04% and a Cohen’s κ of 0.4539 on the BCI IV-2a dataset, and 69.20% and 0.3839, respectively, on the CBCIC dataset, outperforming evaluated baseline methods and demonstrating the potential of the proposed architecture for subject-independent MI-EEG classification.
SwarmPNM combines an adaptive second-moment estimate and positive–negative gradient momentum with attraction towards the lowest-loss weights stored along a single training trajectory. The attraction is inspired by particle swarm optimization, but the method does not maintain a particle population. We specify the update, analyze the attraction locally under Hessian assumptions and obtain a conditional regret bound for an auxiliary strongly convex projected-gradient model with controlled perturbations. Experiments cover 15 analytic test functions, molecular classification with MolCLR-pretrained GIN and GCN encoders on the blood–brain barrier penetration (BBBP), beta-secretase-1 inhibitor (BACE) and human immunodeficiency virus replication-inhibition (HIV) datasets, and nuclear cross-section regression with a multilayer perceptron. SwarmPNM attains the highest mean ROC–AUC in five of six molecular dataset–encoder combinations. In the four-library nuclear comparison, it has the lowest mean absolute error and the highest coefficient of determination; a separate three-seed nuclear experiment under a different protocol favors AdamW. Repeated-run, sensitivity and ablation studies establish that the contribution of attraction depends on the task and tuning, with the learning rate having the largest observed influence in the analytic sensitivity tests. We also report the additional computational cost relative to AdamW. The auxiliary theory does not establish convergence of the complete nonconvex optimizer.
We study learning-augmented combinatorial optimization in the absence of externally supplied labels. We propose CLIO (Contrastive Learning with Iterative Optimization), a search-guided graph neural framework for the maximum independent set (MIS) problem. CLIO is designed around a closed-loop optimization paradigm: graph reduction simplifies the instance, a graph neural network (GNN) produces heuristic node scores, and a local deep search (LDS) procedure enforces feasibility and improves the candidate solution. Critically, the solution returned by LDS is converted into instance-specific pseudo-labels,which are then used to update the GNN through supervised and contrastive objectives. In this way, CLIO learns to solve an unlabeled MIS instance without requiring optimal solutions, external expert decisions, or reward engineering. The contribution of this work is therefore not a single neural architecture, but an integrated neuro-symbolic solving framework that enables label-free adaptation on a per-instance basis. Experiments compare CLIO with greedy and search heuristics, GNN-only and search-only variants, and time-limited solver references. In the ablation study on PPI graphs, the full framework achieved the highest average solution size among the methods except for the variant without graph reduction. LDS contributed the most to solution quality improvement; the closed loop contributed 0.1% and the contrastive loss contributed 0.4% to solution quality improvement, while graph reduction significantly reduced runtime. The current evidence demonstrates that search-generated pseudo-labels provide a viable learning signal for neural combinatorial optimization when no labeled data are available.
