A Knowledge Distillation-Based Approach to Enhance Transparency of Classifier Models

Authors: Yuchen Jiang, Xinyuan Zhao, Yihang Wu, Ahmad Chaddad

AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We selected three public medical data sets (brain tumor, eye disease, and Alzheimer s disease) to test our method. It shows that even when the number of layers is reduced, our model provides a remarkable result in the test set and reduces the time required for the interpretability analysis.
Researcher Affiliation Academia Artificial Intelligence for Personalised Medicine, School of Artificial Intelligence, Guilin University of Electronic Technology Jinji Road, Guilin 541004, China EMAIL
Pseudocode Yes Algorithm 1 provides a basic procedure of our approach.
Open Source Code Yes Code https://github.com/AIPMLab/KD-FMV
Open Datasets Yes We selected three public medical data sets (brain tumor, eye disease, and Alzheimer s disease) to test our method. Brain tumor: This dataset (Nickparvar 2021) is a collection that combines three separate data sets: figshare (Figshare 2017-04-03), the SARTAJ (Bhuvaji 2020), and Br35H (Hamada 2020). Eyes disease: This dataset consists of 4217 Colour fundus photography (CFP)... (Doddi 2023). Alzheimer: The dataset consists of MRI images... (Kaggle 2022).
Dataset Splits Yes Brain tumor: ...our training-to-validation ratio is 8:2. Eyes disease: We randomly divided the images according to training, validation and testing with a ratio of 7:2:1. Alzheimer: The images are randomly divided into training, validation, and testing ratios of 7:2:1. For detailed information on the training/validation and testing sets, refer to Table 1.
Hardware Specification Yes We used an Intel i9-13900k, an NVIDIA Ge Force RTX 4090, and Tensor Flow-gpu 2.6.0 for our simulations.
Software Dependencies Yes We used an Intel i9-13900k, an NVIDIA Ge Force RTX 4090, and Tensor Flow-gpu 2.6.0 for our simulations.
Experiment Setup Yes We used consistent hyperparameter settings, employed the Adam optimizer, set the learning rate to 1 10 4, and configured the batch size to 16.