Distributed Linear Dimensionality Reduction Assisted by Centralized NN for Classification
Keywords: linear dimensionality reduction, edge computing, classfication
Abstract: Linear dimensionality reduction is a widely used technique in data compression, especially under computationally-constrained platforms. This paper presents a linear dimensionality reduction technique tailored for distributed edge devices, balancing resource constraints like data-rate and computing power at the device side, while ensuring high classification accuracy at the server side. The core concept of our approach is the simultaneous training of a unique single-layer for each distributed device, determined by its compression needs, coupled with a centralized deep neural network on the server for all-device classification. A standout feature of our approach is its adaptability: when integrating a new device aiming to compress data in an untrained dimension, only minimal training for the device's initial two layers is needed, leaving the server's centralized deep neural network and the compression layers for all existing devices untouched. Additionally, our findings indicate that the peak accuracy attainable through our method approaches that of the optimal accuracy achievable by the ideal Maximum Likelihood classifier, outperforming traditional matrix decomposition-based techniques like Principal Component Analysis (PCA) and Linear Discriminant Analysis (LDA). Compared to distance-metric-based strategies like Neighborhood Component Analysis (NCA), our technique offers a marked reduction in training complexity for large datasets. Experimental studies show that our approaches result in significant improvements in classification accuracy under the same data-rate requirements compared to existing linear dimensionality reduction approaches on real data sets.
Primary Area: general machine learning (i.e., none of the above)
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 5957
Loading