Machine Learning for Information Exchange and Collaborative Inference over Wireless Channels
Author(s)
Saidutta, Yashas Malur
Advisor(s)
Editor(s)
Collections
Supplementary to:
Permanent Link
Abstract
Deep Learning has revolutionized machine learning and has expanded its reach from image classification to playing games. Not unexpectedly, this has also led to deep learning based design of communication systems. On the other hand, Information Theory provides rich insights in designing such systems. In this thesis, we seek to design deep learning based encoders and decoders for communication and collaborative inference over wireless channels using Information-Theoretic insights. In particular, we look at two main problems, machine learning for signal transmission and machine learning for functional compression and collaborative inference over wireless channels.
In the first part of the thesis, we propose training the encoders and decoders like a Variational Autoencoder for reconstruction over wireless channels. We show that the resulting training minimization objective is an upper bound on the Rate-Distortion optimization objective. This allows us to achieve the state of the art performance over multiple datasets.
In the second part of the thesis, we study the problem of collaborative inference in the IoT setting. Here multiple nodes observe possibly correlated observations, and a central node is interested in computing a function of those observations. Transfer of raw sensory information collected by such devices to cloud servers for processing places a tremendous communication burden on network infrastructure. To alleviate that, we propose three different training schemes of training encoders and decoders and compare their optimization objectives with the objective from the Indirect Rate-Distortion problem. To further lessen the communication burden during training, we propose a training algorithm based on block coordinate descent capable of training these collaborative systems in a distributed manner. This algorithm ensures that the raw sensory information never leaves the collection device during training and inference. Finally, we incorporate explicit privacy constraints and propose novel training schemes for privacy during training and inference.
Sponsor
Date
2022-01-14
Extent
Resource Type
Text
Resource Subtype
Dissertation