On the information theory of clustering, registration, and blockchains

Clustering;Image registration;Universality;Blockchain systems;Scalable system design;Trusted computing;Information-theoretic analysis;error exponents;asymptotic optimality;finite-sample analysis, ,

Raman, Ravi Kiran

关键词: Clustering;    Image registration;    Universality;    Blockchain systems;    Scalable system design;    Trusted computing;    Information-theoretic analysis;    error exponents;    asymptotic optimality;    finite-sample analysis, ,;   

Others  :  https://www.ideals.illinois.edu/bitstream/handle/2142/104833/RAMAN-DISSERTATION-2019.pdf?sequence=1&isAllowed=y

美国|英语

来源: The Illinois Digital Environment for Access to Learning and Scholarship

【 摘 要 】

Progress in data science depends on the collection and storage of large volumes of reliable data, efficient and consistent inference based on this data, and trusting such computations made by untrusted peers. Information theory provides the means to analyze statistical inference algorithms, inspires the design of statistically consistent learning algorithms, and informs the design of large-scale systems for information storage and sharing. In this thesis, we focus on the problems of reliability, universality, integrity, trust, and provenance in data storage, distributed computing, and information processing algorithms and develop technical solutions and mathematical insights using information-theoretic tools. In unsupervised information processing we consider the problems of data clustering and image registration. In particular, we evaluate the performance of the max mutual information method for image registration by studying its error exponent and prove its universal asymptotic optimality. We further extend this to design the max multiinformation method for universal multi-image registration and prove its universal asymptotic optimality. We then evaluate the non-asymptotic performance of image registration to understand the effects of the properties of the image transformations and the channel noise on the algorithms.In data clustering we study the problem of independence clustering of sources using multivariate information functionals. In particular, we define consistent image clustering algorithms using the cluster information, and define a new multivariate information functional called illum information that inspires other independence clustering methods. We also consider the problem of clustering objects based on labels provided by temporary and long-term workers in a crowdsourcing platform. Here we define budget-optimal universal clustering algorithms using distributional identicality and temporal dependence in the responses of workers.For the problem of reliable data storage, we consider the use of blockchain systems, and design secure distributed storage codes to reduce the cost of cold storage of blockchain ledgers. Additionally, we use dynamic zone allocation strategies to enhance the integrity and confidentiality of these systems, and frame optimization problems for designing codes applicable for cloud storage and data insurance. Finally, for the problem of establishing trust in computations over untrusting peer-to-peer networks, we develop a large-scale blockchain system by defining the validation protocols and compression scheme to facilitate an efficient audit of computations that can be shared in a trusted manner across peers over the immutable blockchain ledger. We evaluate the system over some simple synthetic computational experiments and highlights its capacity in identifying anomalous computations and enhancing computational integrity.