A lightweight hybrid deep learning approach for fashion mnist classification with explainable attention visualization

Clinical Snapshot

Key Findings

• P Value: Not reported. No statistical hypothesis testing performed.

• Effect Size: Not formally reported as an effect size. The proposed TinyCNN-LSA model achieves 91.47% accuracy with fewer than 500,000 parameters, compared to VGG-19, ResNet-50, ViT, and Swin Transformer which contain tens of millions of parameters. The accuracy gap versus state-of-the-art Fashion-MNIST models (>96%) represents approximately 4.5 percentage points below top performance.

• Primary Outcome: Classification accuracy of 91.47% on the Fashion-MNIST test set (10,000 images across 10 clothing categories)

• Nnt Or Sensitivity: Not applicable in clinical terms. Model efficiency metric: <500,000 parameters (vs. ~138 million for VGG-19, ~25 million for ResNet-50). CPU-trainable without GPU requirement. No sensitivity/specificity or diagnostic accuracy metrics reported.

• Confidence Interval: Not reported. No confidence intervals or variance estimates provided for any metric.

Clinical Application

From a computer science and engineering perspective, the model's CPU-trainability and low parameter count make it feasible for deployment on edge devices, mobile applications, and low-resource computing environments in the e-commerce sector. However, the 91.47% accuracy on a simplified benchmark may be insufficient for commercial deployment where misclassification has direct revenue implications. This study has no relevance to Australian clinical practice, the Pharmaceutical Benefits Scheme (PBS), Therapeutic Goods Administration (TGA) regulatory frameworks, or Royal Australian College of General Practitioners (RACGP) guidelines. If the XAI methodology were adapted for medical image analysis (e.g., dermatology, radiology), TGA oversight of AI-based medical devices under the Software as a Medical Device (SaMD) framework would become relevant. However, no such application is proposed or validated in this study. This study has no direct clinical applicability. It addresses automated fashion image classification for e-commerce product categorisation. There is no patient population, clinical condition, diagnostic question, or therapeutic intervention involved. The MeSH term 'Humans' assigned to this study appears to be a cataloguing error.

Abstract

The classification of fashion images is an essential task in the e-commerce sector, where accurate categorization improves user experience and refines product discovery. Convolutional Neural Networks (CNNs) and Transformers have demonstrated strong performance in image classification tasks due to their ability to learn complex visual features. However, deep variants of these architectures, such as VGG-19, ResNet-50, Vision Transformer (ViT), and Swin Transformer, contain tens of millions of parameters, requiring high memory and powerful GPUs for training, which makes them less suitable for low-resource and edge device environments. To address these limitations, this research proposes a lightweight hybrid architecture, TinyCNN with Linear Self-Attention (LSA), optimized for resource-constrained settings. The proposed model contains fewer than half a million parameters and is trainable on a CPU, achieving a classification accuracy of 91.47% on the Fashion-MNIST dataset. In addition, multiple Explainable Artificial Intelligence (XAI) techniques are implemented, including Self-Attention visualization, Multi-Head Attention, Attention Flow, Attention Rollout, Fixed query position attention maps, Integrated Gradients, LIME, and SHAP, to provide visual interpretability of the model's predictions and enhance transparency in its decision-making process.

References

1. 1.Ahmad, H., Lashari, T. A., Lashari, S. A., Khan, I., & Jabeen, F. (2026). A lightweight hybrid deep learning approach for fashion mnist classification with explainable attention visualization. PLOS ONE. https://doi.org/10.1371/journal.pone.0351671