A dataset of deep learning performance from cross-base data encoding on MNIST and MNIST-C

Area of Science:

• Machine Learning
• Deep Learning
• Data Science

Background:

• Effective data representation is crucial for machine learning (ML) and deep learning (DL) algorithms to accurately capture patterns and make reliable predictions.
• While data preprocessing is widely documented, novel methods for data representation to enhance ML model performance are underrepresented in scientific literature.

Purpose of the Study:

• To compile a dataset of convolutional neural network (CNN) model performance metrics.
• To evaluate CNN performance when trained and tested on diverse numerical base representations of the MNIST and MNIST-C datasets.
• To provide a resource for the research community to analyze trends between data encoding and ML model performance.

Main Methods:

• The study involved training and testing CNN models on datasets derived from MNIST and MNIST-C.
• Data was encoded using a range of numerical base representations.
• Model performance was systematically recorded for each data representation.

Main Results:

• The dataset captures CNN performance variations across different numerical bases for MNIST and MNIST-C data.
• This compilation allows for the investigation of how data encoding strategies influence model accuracy and generalization.
• The findings highlight the sensitivity of CNNs to the numerical representation of input data.

Conclusions:

• The generated dataset serves as a valuable resource for understanding the impact of data representation on ML/DL model performance.
• Further research can leverage this dataset to explore cross-base data encoding for improved performance in various real-world applications.
• This work encourages the development of novel data representation techniques to advance ML and DL capabilities.