Adaptive Dimensionality Reduction for Efficient Deep Learning on Temporal Datasets

Abstract In the domain of computer vision and its related research fields, especially in temporal data handling, the analysis and processing of high dimensional data is a huge problem. Challenges in deep learning-based high dimension data processing often arise from their variable length sequences and their complex data patterns. As the use of high dimensional data continues to grow, the demand for dimension reduction of complex data is growing rapidly. Dimensionality reduction is the process of reducing a high-dimensional matrix of a dataset into a lower dimension. Standard approaches based on principal component analysis, independent component analysis and others, including simple padding and truncation, can lead to data loss or unwanted computational overhead. To address this issue, we introduce a VidSqeOpt method for dimensionality reduction based on feature extraction which exploits statistical approaches with reduced dimension and optimal length finding. Our method is validated on several benchmark datasets that are based on temporal data with complex dimensionality. Cross comparison with the baseline methods shows that our proposed approach adapts to diverse tasks, achieving up to 2.0% higher accuracy while reducing per-frame processing time by at least 23.7%, down to ≤0.45 seconds.