Leveraging Perceiver IO for Efficient Compression of Temporal Signals A Memory Constrained Approach
Year: 2024
Type: conference-paper
Abstract: The compression of temporal signals presents significant challenges in the era of big data, where efficient extraction of valuable information from large datasets is critical across various applications.This paper investigates the application of the Perceiver IO model, a novel deep learning architecture designed for flexible multi-modal input processing, in compressing temporal signals.Unlike traditional compression methods, which struggle with variable signal lengths and complex dependencies, the Perceiver IO utilizes an attention-based mechanism that dynamically captures both shortterm and long-term patterns in data.In our study, we emphasize the achieved compression ratios and the effective extraction of relevant information over reconstruction quality.Empirical evaluations demonstrate that the Perceiver IO model can achieve a compression ratio of approximately 2500:1 with a maximum Mean Squared Error (MSE) of 0.01, showcasing its capability to retain essential signal features.We underscore the potential of the Perceiver IO to advance temporal signal compression, paving the way for future studies into efficient data representation via deep learning frameworks.
Institution: Université de Technologie de Troyes
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W4411936797