10a.002.TAU_WP6 - Operational Support Estimator Network (OSEN) for Sparse Support Estimation and Learning-Aided Compressive Sensing

Project - Summary

Support estimation (SE) of a sparse signal refers to finding the location indices of nonzero elements in a sparse representation. In general, traditional SE approaches are based on iterative algorithms consisting of greedy methods and optimization techniques. This is due to the fact that a vast majority of them apply sparse signal recovery (SR) techniques to obtain support sets. On the other hand, the SE task is less complicated than SR and computing direct mapping from the measurement (for example, compressively sensed) to the nonzero locations is more desired. This project proposes a novel approach for learning such a mapping from the measurement signal. To accomplish this objective, the Operational Sparse Support Estimator Networks (OSENs), each with a compact configuration, are designed. The proposed network composes of 2D-operational layers with the non-linear neuron model. Hence, the learning capability of neurons (filters) is greatly improved with shallow architectures compared to the traditional convolutional layers performing linear transformation. The proposed OSEN can be a crucial tool for the following scenarios: 1) real-time and low-cost SE can be applied in any mobile and low-power edge device for anomaly localization, simultaneous face recognition, and so on and 2) OSEN’s output can directly be used as “prior information,” which improves the performance of sparse SR algorithms. The results over the benchmark datasets show that state-of-the-art performance levels can be achieved by the proposed approach with a significantly reduced computational complexity thanks to proposed direct SE approach and operational layers.

Project - Team

Team Member Role Email Phone Number Academic Site/IAB
Moncef Gabbouj PI 358 40 073 6613 Tampere University
Serkan Kiranyaz Co-PI 97 43 063 5600 Tampere University
Mete Ahishali Researcher 358 46 552 3736 Tampere University
Matti Vakkuri Project Mentor N/A Haltian
Tomi Teikko Project Mentor N/A N/A Haltian
Gunnar Hansen Project Mentor N/A N/A Haltian

Project - Novelty of Approach

The existing SE studies are based on first applying SR and then performing threshold over the reconstructed signal. However, the SR task is more challenging than the SE and in many cases such as classification and anomaly detection, the SE would be sufficient. Thus, our proposed approach is trained to produce direct support sets from the measurements without performing SR. In this way, we achieve the improved accuracy and computational efficiency. In the previous project, the CSENs were proposed equipped with traditional convolutional layers. The OSENs with operational layers have superior learning ability with non-linear kernel transformation functions compared to the CSENs. Furthermore, the proposed super-neuron model has non-localized kernels so that the kernel locations are not fixed and they are optimized during the training of the network. It is observed that the proposed OSEN have improved the classification and SE accuracy of the CSENs with more compact network structures.

Project - Deliverables

1 Improved classification framework with the hybrid loss penalizing both SE estimation and classification errors during training.
2 Learning-aided CS scheme with the prior information produced by the OSEN.

Project - Benefits to IAB

Advanced Machine Learning especially for scarce data is a great asset for any service company offering AI solutions to real world problems.

The proposed approach can be used in many tasks such as classification, anomaly localization, and various CS based applications including compressive video surveillance systems and Distributed Compressive Sensing based surveillance.

A hybrid method can still be used with the proposed approach: the output of OSEN also provides the probability of support sets that can be used as prior in the classical CS schemes. In this way, the newly introduced learning-aided CS scheme with OSENs addresses the time complexity issue of the recovery algorithms.

The proposed Machine Learning techniques, designed especially for scarce data can be used in many tasks and various applications including (distributed) video surveillance systems empowers Haltian and other IAB in the intelligent built environment field to enhance their EB solution portfolio and help them improve their products.

Project - Documents

projects/year10/10a.002.tau_wp6.txt · Last modified: 2022/11/22 09:11 by sally.johnson