• HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
Tuesday, November 17, 2026
BIOENGINEER.ORG
No Result
View All Result
  • Login
  • HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
        • Lecturer
        • PhD Studentship
        • Postdoc
        • Research Assistant
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
  • HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
        • Lecturer
        • PhD Studentship
        • Postdoc
        • Research Assistant
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
No Result
View All Result
Bioengineer.org
No Result
View All Result
Home NEWS Science News Technology

Deep Learning Architecture Shows High Uncertainty on Physical Education Student Dataset

by
November 12, 2026
in Technology
Reading Time: 6 mins read
0
Deep Learning Architecture Shows High Uncertainty on Physical Education Student Dataset

Deep Learning Architecture Shows High Uncertainty on Physical Education Student Dataset

Share on FacebookShare on TwitterShare on LinkedinShare on RedditShare on Telegram

A multi-branch deep learning framework designed to evaluate physical education students achieved a Macro-F1 classification score of 0.545 across five-fold cross-validation on a public benchmark dataset, but it failed to outperform conventional gradient-boosted decision trees and random forests. Developed by researchers Yunlai Lv and Siyuan Li, the neural network architecture named MartialSense-Net was formulated to address persistent challenges in quantitative physical education assessment, where conventional evaluation often depends on experiential instructor grading or isolated fitness tests. The computational study, published in Discover Artificial Intelligence, introduced an end-to-end framework intended to model complex interactions among distinct categories of physical education data by grouping heterogeneous tabular features into dedicated semantic branches before merging them through an attention mechanism. However, comprehensive empirical testing revealed that the neural framework exhibited substantial fold-to-fold variance, trailed standard machine learning baselines in predictive consistency, and produced negative coefficients of determination on continuous score regression tasks.

Traditional instructional evaluation in university courses such as Wushu, or Chinese martial arts, emphasizes a multi-dimensional balance encompassing movement standardization, force expression coordination, theoretical knowledge, and classroom conduct. Instructors typically struggle to continuously and quantitatively monitor student trajectories across these disparate domains using subjective visual scoring or isolated final examinations, frequently missing opportunities to deliver timely pedagogical interventions for lagging learners. To conceptualize an automated alternative, the authors hypothesized that structured deep learning combined with multi-head self-attention could capture the statistical dependencies among diverse student characteristics. Rather than evaluating the system on dedicated martial arts motion-tracking data, the researchers tested the architecture as a proof of concept on a publicly accessible benchmark dataset titled student_pe_performance.csv. Because the underlying benchmark contained standard physical education variables rather than authentic martial arts measurements of posture, rhythm, or routine accuracy, the experimental design tested the structural capacity of the network rather than confirming sport-specific clinical or educational validity.

The benchmark dataset comprised 500 individual student records across 20 source variables spanning demographic attributes, physical fitness measures, theoretical knowledge examinations, behavioral characteristics, and course performance outcomes. To construct the model inputs, the researchers excluded student identification codes and target outcome labels, retaining 16 original predictor variables that were normalized and encoded. Within each cross-validation fold, numerical values were standardized through Z-score transformations derived strictly from training partitions to prevent data leakage, while categorical variables underwent one-hot encoding. Following preprocessing, the encoded input vectors comprised 30 dimensions partitioned across four distinct semantic branches: a five-dimensional Skill branch, a six-dimensional Physical fitness branch, a six-dimensional Theory branch, and a 13-dimensional Behavior and context branch. The researchers noted that this structural grouping operated as an architectural prior rather than an empirically validated latent educational construct, particularly because demographic indicators like age, grade level, and gender were grouped into the behavioral branch as contextual controls.

At the core of the MartialSense-Net framework, a Multi-Branch Feature Extraction module independently processes each of the four feature partitions using separate linear projection layers combined with rectified linear unit activations. This initial step maps the heterogeneous inputs of each domain into uniform 16-dimensional latent representation vectors, creating a sequence of four branch tokens. These tokens are subsequently ingested by a Cross Feature Attention Fusion module equipped with four-head self-attention. The attention mechanism calculates query, key, and value matrices across the token sequence, producing scaled dot-product attention distributions that model pairwise statistical interactions among the semantic branches before adding a residual connection back to the original representations. The resulting fused 64-dimensional sequence is flattened and fed into a three-layer multilayer perceptron encoder with layer dimensions of 32, 64, and 32 units, all utilizing rectified linear units. The entire architecture totals 8,036 trainable parameters and terminates in two multi-task prediction heads: a classification head that categorizes students into low, average, or high performance levels, and a regression head that predicts overall physical education examination scores.

To train the multi-task architecture simultaneously on discrete and continuous outcomes, the investigators formulated a combined objective function pairing class-weighted cross-entropy loss with mean squared error regression loss. The continuous regression target was standardized prior to training and weighted by a scaling factor set to 0.5, while class weights in the cross-entropy term were scaled inversely proportional to class frequencies to counteract severe label imbalance. Model parameters were updated using the Adam optimizer with a learning rate of 0.001, a weight decay coefficient of 0.0001, and a mini-batch size of 32 over 80 training epochs. Analysis of the benchmark outcome targets revealed that the three performance categories exhibited moderate, statistically significant correlations with continuous performance scores, yielding a Spearman correlation coefficient of 0.347 and a Kruskal-Wallis metric of 69.660. However, the score distributions among the categories overlapped extensively, demonstrating that the categorical performance label was not a simple deterministic thresholding of the continuous score.

Rigorous evaluation using stratified five-fold cross-validation demonstrated that MartialSense-Net achieved a mean Macro-F1 score of 0.545, surpassing a basic multilayer perceptron baseline and an architectural variant lacking multi-branch extraction, which attained a Macro-F1 of 0.483. Despite these architectural gains, the proposed deep network did not exceed the performance of traditional tree-based ensemble models evaluated on identical data partitions. A standard Random Forest classifier secured the highest overall Macro-F1 score at 0.583, while gradient-boosted algorithms CatBoost and XGBoost also demonstrated superior predictive discrimination. Supervised tabular neural baselines, such as SAINT and FT-Transformer, likewise exhibited competitive classification performance. The authors highlighted that on small-scale tabular datasets characterized by high heterogeneity, the theoretical representation-learning advantages of deep neural networks remain constrained, allowing conventional non-parametric decision tree algorithms to retain competitive accuracy.

To investigate whether the primary five-fold cross-validation results reflected a single favorable data split or persistent performance advantages, the investigators executed a repeated stratified five-fold protocol consisting of 25 matched held-out evaluations across the leading models. In this extended test, MartialSense-Net demonstrated substantially greater metric dispersion than the tree-based ensembles, recording an average Macro-F1 that trailed Random Forest by 0.015 and XGBoost by 0.011. While paired Wilcoxon tests across the 25 matched folds indicated no statistically significant difference in median performance, bootstrap 95% confidence intervals for the paired differences were exceptionally wide, spanning from negative 0.044 to positive 0.017 relative to Random Forest and negative 0.039 to positive 0.020 relative to XGBoost. These wide confidence bounds confirmed that the dataset sample size was insufficient to establish either the statistical superiority or the functional equivalence of the deep learning architecture.

Ablation experiments targeting specific network components yielded critical insights into the internal mechanisms of MartialSense-Net. While removing the Multi-Branch Feature Extraction module caused Macro-F1 performance to drop sharply from 0.545 to 0.483, entirely removing the Cross Feature Attention Fusion attention mechanism yielded a slightly higher Macro-F1 of 0.550. Consequently, while modular multi-branch feature encoding provided clear structural benefits, the addition of complex multi-head self-attention failed to produce measurable mean improvements. Furthermore, inspection of the learned cross-branch attention weights revealed that the fold-averaged attention matrix was nearly uniform, with individual branch weights ranging narrowly between 0.205 and 0.336. The average attention weights assigned to the Skill, Physical, Theory, and Behavior keys were 0.216, 0.258, 0.216, and 0.310, respectively, demonstrating that the model did not converge on a distinct or sparse hierarchical relationship among instructional dimensions.

The regression evaluation exposed substantial modeling limitations, as MartialSense-Net and all evaluated benchmark regressors failed to achieve positive coefficients of determination on the continuous score target. The primary 80-epoch multi-task model suffered from severe overfitting, registering an average test mean absolute error of 10.049 and a negative coefficient of determination. Even a dedicated regression-only neural model trained with inner-loop early stopping achieved a coefficient of determination of negative 0.030, failing to outperform a simple baseline predicting the training-set mean. Learning-curve analyses showed that although the regression model improved when training partitions increased from 160 to 320 records, expanding training data further to 400 records produced no additional performance gain. The persistent failure across both linear and non-linear regressors indicated that the available tabular predictors contained limited predictive information regarding the continuous examination outcome.

The study also analyzed extreme class imbalance as a principal driver of classification instability, given that high-performing students constituted only 12 out of 500 total records, or 2.4% of the cohort. Under standard weighted cross-entropy training, the model correctly identified only two out of 12 high performers in pooled out-of-fold predictions. Alternative imbalance mitigation strategies implemented strictly within training partitions, such as focal loss and class-balanced mini-batch sampling, increased minority recall but caused catastrophic reductions in overall Macro-F1 discrimination. Furthermore, the presence of only two or three high performers in each five-fold validation split meant that a single misclassification dramatically shifted fold-level Macro-F1 scores, accounting for the elevated standard deviation of 0.091 observed in the full model. The authors concluded that validating machine learning systems for physical education requires authentic longitudinal datasets containing direct movement kinematics, standardized force measurements, and rigorous external validation across institutions rather than relying on small general physical education benchmarks.

Subject of Research: Evaluation of multi-branch deep learning architectures for physical education student outcome prediction

Article Title: Deep learning for learning outcome prediction and evaluation in Wushu physical education

Article References: Lv, Y., & Li, S. (2026). Deep learning for learning outcome prediction and evaluation in Wushu physical education. Discover Artificial Intelligence, 6(1), Article 1853. https://doi.org/10.1007/s44163-026-02693-z

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02693-z

Keywords: deep learning, physical education, tabular data, machine learning baselines, random forest, XGBoost, multi-task learning, educational data mining, cross feature attention, class imbalance, stratified cross-validation, model evaluation, academic performance, feature extraction, performance prediction, model ablation

News Source: Blake Davidson. (November 12, 2026). Deep Learning Architecture Shows High Uncertainty on Physical Education Student Dataset. Scienmag.

Tags: academic performanceclass imbalancecross feature attentiondeep learningeducational data miningfeature extractionmachine learning baselinesmodel ablationModel EvaluationMulti-task learningPerformance predictionphysical educationRandom ForestStratified cross-validationtabular dataXGBoost
Share12Tweet7Share2ShareShareShare1

Related Posts

Guiding Wave-Like Fisher Information Flow Enhances Nanophotonic Optical Metrology Precision

Guiding Wave-Like Fisher Information Flow Enhances Nanophotonic Optical Metrology Precision

November 17, 2026
New Proximal Method Generates Unique Fractal Attractors in Neutrosophic Metric Spaces

New Proximal Method Generates Unique Fractal Attractors in Neutrosophic Metric Spaces

November 17, 2026

Temporal Hybrid Neural Network Achieves Over 99 Percent DDoS Detection Accuracy

November 17, 2026

Presence Dynamics Framework Proposes How Multi-Stage Media Experiences Reconfigure Immersion

November 17, 2026

POPULAR NEWS

  • Mobile Health Units With Community Workers Address Critical Hypertension Care Gaps

    Mobile Health Units With Community Workers Address Critical Hypertension Care Gaps

    29 shares
    Share 12 Tweet 7
  • Brazilian EQ-5D Mapping Algorithm Offers Alternative to European-Based Standards

    29 shares
    Share 12 Tweet 7
  • Self-Assembled Berberine-Gallic Acid Nanorods Ameliorate Ulcerative Colitis in Mice

    29 shares
    Share 12 Tweet 7
  • Tablet Augmented Reality System May Reduce Preoperative Anxiety in Pediatric Surgical Patients

    29 shares
    Share 12 Tweet 7

About

We bring you the latest biotechnology news from best research centers and universities around the world. Check our website.

Follow us

Recent News

Inhibiting EID1 Boosts Immune Cell Cooperation to Suppress Experimental Metastasis

Global Research Consortium Analyzes Disorganized Speech Patterns to Measure Psychosis

Guiding Wave-Like Fisher Information Flow Enhances Nanophotonic Optical Metrology Precision

Subscribe to Blog via Email

Enter your email address to subscribe to this blog and receive notifications of new posts by email.

Join 88 other subscribers
  • Contact Us

Bioengineer.org © Copyright 2023 All Rights Reserved.

Welcome Back!

Login to your account below

Forgotten Password?

Retrieve your password

Please enter your username or email address to reset your password.

Log In
No Result
View All Result
  • Homepages
    • Home Page 1
    • Home Page 2
  • News
  • National
  • Business
  • Health
  • Lifestyle
  • Science

Bioengineer.org © Copyright 2023 All Rights Reserved.