Browsing by Subject "Structure Learning"

Now showing 1 - 1 of 1
  • Thumbnail Image
    Item
    Information Processing: Thermodynamics of Computations and Network Inference
    (2018-12) Talukdar, Saurav
    The present times is seeing a surge in computing due to several applications like smart grid, autonomous vehicles, social networks and many more directly touching our lives. A key enabler is the technological advancements in information storage and information processing technologies, for example, data centers, wireless communication, cloud computing, artificial intelligence to name a few. A present day iPhone is more powerful than the computer used by NASA in the Apollo mission as well as the 1997 IBM super computer which defeated Grand-master Gary Kasparov. The consequence is that data centers are now projected to consume about 20% of Earth's power by 2030. In the first part of this dissertation, fundamental computation mechanisms and their energy consumption are explored. Erasure or reset of information stored in a single bit memory is studied in detail. In particular, we experimentally demonstrate erasure using almost the minimum possible amount of energy required for the erasure of a bit of information, as dictated by the Landauer's principle. Optical traps are used to achieve this and a detailed modeling of the dynamics in optical traps is discussed first, which is used to develop a Monte Carlo simulation to study physics of information erasure. Finally, we analyze erasure of information using a Gaussian mixture approach and conclude with the trade-off between reliability of information erasure, minimum energy consumption in information erasure and size of the memory bit. The second part of this dissertation is focused on information processing from diverse sources with the goal of enabling decision making for energy efficiency, safe operation, human comfort and minimize costs in complex physical systems. For example: the concept of smart home uses information about weather, energy state of the grid, usage pattern of the home owner etc to decide on control set points of thermostats or air conditioners or heaters. In such complex systems, it is often not clear which entities influence which other entities and is nearly impossible to develop a first principles model. Here, the focus is on developing algorithms with guarantees for inference of presence or absence of relationships among observed entities in a complex system. The framework of network representation of complex systems of linear dynamical systems with wide sense stationary forcing is used and algorithms which infer the relationships amongst the observed entities using only time series observations without any knowledge of system parameters is developed. The performance of the algorithms is shown on simulation examples from power grid, building thermal dynamics and experimental realization of a wireless network of agents. Finally, extensions of the algorithms in the case of unobserved states and non stationary processes is discussed.