Data aware sparse non-negative signal processing
View/ Open
Date
14/06/2022Author
Voulgaris, Konstantinos
Metadata
Abstract
Greedy techniques are a well established framework aiming to reconstruct signals which
are sparse in some domain of representations. They are renowned for their relatively low
computational cost, that makes them appealing from the perspective of real time applications. Within the current work we focus on the explicit case of sparse non–negative
signals that finds applications in several aspects of daily life e.g., food analysis, hazardous materials detection etc. The conventional approach to deploy this type of algorithms does not employ benefits from properties that characterise natural data, such
as lower dimensional representations, underlying structures. Motivated by these properties of data we are aiming to incorporate methodologies within the domain of greedy
techniques that will boost their performance in terms of: 1) computational efficiency
and 2) signal recovery improvement (for the remainder of the thesis we will use the
term acceleration when referring to the first goal and robustness when we are referring
to the second goal). These benefits can be exploited via data aware methodologies that
arise, from the Machine Learning and Deep Learning community.
Within the current work we are aiming to establish a link among conventional
sparse non–negative signal decomposition frameworks that rely on greedy techniques
and data aware methodologies. We have explained the connection among data aware
methodologies and the challenges associated with the sparse non–negative signal decompositions: 1) acceleration and 2) robustness. We have also introduced the standard data
aware methodologies, which are relevant to our problem, and the theoretical properties
they have. The practical implementations of the proposed frameworks are provided
here. The main findings of the current work can be summarised as follows:
• We introduce novel algorithms, theory for the Nearest Neighbor problem.
• We accelerate a greedy algorithm for sparse non–negative signal decomposition
by incorporating our algorithms within its structure.
• We introduce a novel reformulation of greedy techniques from the perspective of
a Deep Neural Network that boosts the robustness of greedy techniques.
• We introduce the theoretical framework that fingerprints the conditions that lay
down the soil for the exact recovery of the signal.