Content area
Machine learning is a fundamental tool that is incorporated in every field across academia and other industries. Due to the large amount of data needed for training machine learning models, lossy compression plays a crucial role in storing data. Machine learning involves the use of algorithms and models to learn patterns in data. This allows the AI to make decisions without specific programming. On the other hand, compression utilizes encoding and decoding techniques to reduce the size of files. Compression is either lossy or lossless, lossy causes a loss of data while lossless preserves the data.
This dissertation will explore the performance of machine learning when working with data that has undergone lossy compression. The performance metrics that are being studied looks at how accurate the model’s inference will perform (i.e. accuracy, intersection over union) depending on the task. The issues with machine learning performances on lossy data involve the following: data storage, data transfer bandwidth, and processing on the intersection between machine learning and lossy compression. Over these various tasks, machine learning in different domains will be examined to investigate how meaningful patterns in the distorted data is extracted.
One approach explored in this work involves the analysis and design of various neural network models allowing the research to manage lossy compressed data in an isolated format. The primary focus will be on machine learning that works with image data. This also includes finding implications across various domains of image processing. Examples of this are object detection, semantic segmentation, and image classification. Balancing the compression ratio and the data quality is critical to measure performance of the model in compliance with the space used.