August 29 · Preprint · arXiv · Cognitive Neuroscience
Single-trial EEG analyses are often organized around events and latencies, yet EEG-based reaction-time (RT) prediction is posed as scalar regression on a fixed stimulus-locked window. RT is treated as a window-level label rather than timing evidence about response-relevant dynamics. Here we reformulate trial-wise RT decoding as event-time posterior modeling. Instead of predicting RT directly, the model estimates a posterior over response-relevant event times, $p(t_{\mathrm{event}}\mid X)$, and uses its mean as the RT estimate. This treats behavioral latency as a weak observation of latent response-relevant timing. We evaluate this formulation on the Healthy Brain Network contrast change detection EEG task under a subject-disjoint, release-separated protocol. Across five seeds, distributional event-time supervision consistently improves held-out RT prediction relative to scalar regression and temporal-readout controls. Controlled objective comparisons isolate supervision of the event-time distribution, rather than expectation-based readout alone, as the source of this gain. Architecture controls show that the effect persists across four temporal backbones and is not explained by model scale. Beyond point prediction, posterior geometry characterizes concentration, target alignment, and interval behavior, while observation-noise calibration separates latent concentration from predictive uncertainty over RT. Shifted-crop inference probes shortcut use versus temporal localization. Matched shift-jitter improves robustness, increases mean sensitivity, and moves predictions more often in the expected crop-relative direction. Sensitivity remains below ideal crop-relative localization, leaving a clear equivariance gap. Together, these results establish event-time posterior modeling as a probabilistic and interpretable formulation for linking single-trial EEG dynamics to behavioral timing.
Anuar AimoldinOpen access
August 31 · Article · Вестник КазАТК · Political Science and International Relations
Analysing a country's body of law by computer is harder than it first looks. The texts sit in a deep hierarchy, point at each other constantly, change often, and come with few tools tuned to the local language. This paper puts together an open pipeline of NLP and machine-learning parts and runs it on the five main codes of Kazakhstan: the Criminal Code, the Code of Administrative Offenses, the Civil Code, the Labor Code, and the Tax Code. In its current-version form the corpus holds 3,062 articles and 631,280 tokens. A second, historical corpus of 563 versions from 1996 to 2025 lets us watch the law change over time. The pipeline cleans text with spaCy, builds TF-IDF features, and trains four supervised classifiers. It also fits LDA topic models, counts bigrams, embeds every article with a multilingual sentence model (E5-base, 768 dimensions), projects those vectors with UMAP, and groups them with HDBSCAN. A grammar-aware parser reads penalty values out of the text, even when the numbers are written as Russian words, and a version-by-version diff tags what happened to each article. A linear SVM over TF-IDF separates the five codes with 98.48 % accuracy; a Random Forest reaches 94.92 % on the three-way penalty-type task. HDBSCAN finds 12 coherent clusters at 91.5 % average purity, and parallel provisions in the Criminal and Administrative codes come out as close as 0.99 by cosine similarity. The citation graph has 6,436 edges, and 92.0 % of them stay inside a single code, so whatever ties the codes together is not written into the text itself. Penalty parsing returns 4,524 records from the current codes and 685 transitions from the historical ones, and the two pull in opposite directions: prison terms and corporate fines grow harsher while penalties aimed at ordinary citizens’ ease. Of the 8,225 edits we looked at, 86 % are cosmetic or minor. The code, the notebooks, and the intermediate files are all public, and the same pipeline should carry over to other hierarchical corpora that use structured citations.
BMBobur MukhsimbayevOpen access
August 31 · Article · Вестник КазАТК · Automotive Engineering
The comparison of trajectory prediction models is complicated by the fact that architectural differences are coupled with variations in input representation and training procedures. This study introduces a partially controlled evaluation protocol in which datasets, temporal task configuration, evaluation metrics, and computational constraints are fixed, while preserving architecture-specific input representations and individualized model tuning. In contrast to conventional approaches, the analysis is conducted not only on aggregated metric values but also at the level of scene-wise error distributions. Experiments on the ETH/UCY, Argoverse 2, and Waymo Open Motion datasets demonstrate that differences between models in terms of mean ADE and FDE remain limited and are accompanied by comparable intra-model variability. Statistical testing reveals significant differences (Friedman test, p < 0.001); however, the effect size remains modest (Kendall’s W ≈ 0.17–0.19), indicating weak agreement in model rankings. The findings suggest that the interpretation of model comparisons is strongly dependent on the experimental setup and the level of metric aggregation. The proposed protocol enables a clearer separation between the influence of experimental conditions and intrinsic model properties, thereby complementing—rather than replacing—existing benchmarking practices.
ZUZhanat UmarovaOpen access
