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

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Fashion image classification is crucial for e-commerce, but deep learning models like CNNs and Transformers are computationally expensive.
• High parameter counts in models like VGG-19, ResNet-50, ViT, and Swin Transformer limit their use in low-resource and edge devices.

Purpose of the Study:

• To develop a lightweight hybrid architecture for fashion image classification suitable for resource-constrained environments.
• To achieve high classification accuracy with a model that requires minimal computational resources.

Main Methods:

• Proposed a novel lightweight hybrid architecture: TinyCNN with Linear Self-Attention (LSA).
• The TinyCNN-LSA model features fewer than half a million parameters, enabling CPU training.
• Implemented various Explainable Artificial Intelligence (XAI) techniques for model interpretability.

Main Results:

• The TinyCNN-LSA model achieved 91.47% classification accuracy on the Fashion-MNIST dataset.
• The model is trainable on a CPU, demonstrating its efficiency for low-resource settings.
• XAI techniques provided visual interpretability of the model's prediction decisions.

Conclusions:

• The TinyCNN-LSA architecture offers an effective and efficient solution for fashion image classification in resource-constrained environments.
• The model's lightweight nature and high accuracy, combined with XAI for transparency, make it suitable for practical e-commerce applications.
• This research contributes to making advanced AI capabilities more accessible on edge devices.