Deep Learning Performance Dataset for MNIST and MNIST-C
Author Information
Author(s): Lawrence McKnight, Chandra Jaiswal, Issa AlHmoud, Balakrishna Gokaraju
Primary Institution: North Carolina A&T State University
Hypothesis
Does cross-base data encoding enhance the performance of deep learning models on MNIST and MNIST-C datasets?
Conclusion
The dataset provides insights into how different numerical base representations affect model performance in image classification tasks.
Supporting Evidence
- The dataset allows practitioners to identify statistically significant train-test base pairs quickly.
- It serves as a resource for understanding the relationship between numerical base representations and model prediction accuracy.
- The dataset format can be adapted to various machine learning models and problem sets.
Takeaway
This study created a dataset to help researchers see how changing the number system used for data can make machine learning models better at recognizing handwritten numbers.
Methodology
The dataset includes class-wise accuracy data from convolutional neural networks trained on MNIST and MNIST-C datasets encoded in bases 2 through 10.
Limitations
The dataset is limited to a single CNN architecture and does not include additional performance metrics such as training time or computational complexity.
Digital Object Identifier (DOI)
Want to read the original?
Access the complete publication on the publisher's website