Perceptual Fidelity in Learned Image and Video Processing

Perceptual quality of images or videos after compression plays an important role in designing image or video processors. If the system performs a lossy compression, the so-called rate-distortion theory has been developed to characterize the tradeoff between the compression rate and distortion of the reconstruction from the input. However, through experiments, it has been observed that satisfying a small distortion does not necessarily lead to a high perceptual quality in the reconstruction. So, the perception metric has been introduced into the system evaluation parameters which basically measures the divergence between distributions of the input and reconstruction. Our project aims at investigating the perceptual quality in different image and video compression systems.

Distributed Testing and Learning in Communication Systems

Recently, there has been a significant improvement in the ability to process and store large amounts of data which will change the design of future communication networks and machine learning systems, dramatically. From the distributed machine learning side, a new data paradigm is emerging. The conventional distributed machine learning systems dealt with only a few users each having access to “big” datasets. In the new paradigm which is commonly referred to as “federated learning”, there are massive number of users that each of them has a “small” dataset. For these massively distributed systems to work efficiently, data compression is critical since it reduces the communication load. From the networks perspective, by the decreasing costs of sensors, several new distributed monitoring and alert systems emerge as part of the Internet of Things (IoT). These systems consist of distributed sensor nodes that communicate information about their measured signals to one or multiple decision centers that have to decide on some hypotheses. A simple approach for designing such systems is that the sensors compress their signals and send these compressed versions to the decision centers. This approach however is not feasible when the sensor nodes have stringent power constraints or when the networks saturate because of the large number of transmitting devices. Our project aims at designing learning and testing algorithms for these communication systems.

Security and Privacy Algorithms in IoT

The IoT has several unique characteristics that distinguish it from the existing network architectures. The main characteristics are the deployment of massive number of low-complexity terminals, the need for light or no infrastructure and basic applications of data gathering and inference. These characteristics shape the concept of security in IoT where some new fundamental insights should be proposed.