Abstract

While designing computer applications to solve real world problems, engineers and Machine Learning researchers have the freedom to design models specific to the situation at hand. If the application is required to have high latency or low memory usage, the model is typically tweaked at the design stage itself to incorporate these properties, and such models cannot be used in cases where reliability or explainability takes precedence, even if the underlying task is the same. However, in the case of cognitive agents, the strategies and algorithms used to make decisions need to adapt to the situation in an online manner. A popular idea in cognitive science used to explain how animals make important tradeoffs is that of viewing Decision Making as an Evidence Accumulation process. By modelling Decision Making as sampling evidence until a threshold, it is possible for such models to showcase different behaviours as per the needs of the situation. In this thesis, we take the sequential sampling approach and combine it with Reinforcement Learning frameworks in an attempt to move towards a more comprehensive model of Decision Making. In the first portion of the thesis, we explore how Linear Ballistic Accumulators can be incorporated as an action selection mechanism into Q-Learning models, which we refer to as RLLBAs. One important advantage this brings about is the ability of RLLBAs to utilise reaction time data in addition to choice data. We compare the performance of RLLBAs and conventional models in a non-trivial Grid Navigation Task with three action choices. It was found that RLLBAs were able to predict the actions taken by the subjects as well as conventional RL models while at the same time providing good predictions of the reaction time data. In addition, it was also shown that RLLBAs show significant differences in the goodness-of-fit between various forms of arbitration between Model-Free and Model-Based RL, something which is typically harder to achieve with choice data alone. In the second portion of the thesis, we explore how evidence accumulation can be realized in RL neural networks. For this purpose, we take inspiration from existing literature on anytime neural networks and structured reservoirs. The central idea here is to structure the connections of the reservoir so that activity in the network propagates forward across time. As the activity propagates forward it undergoes more processing and becomes less noisy, meanwhile, as the output layer has access to earlier parts of the reservoir, the model can still respond quickly to sudden changes in the environment if relevant. On further experimenting with connectivity patterns found in the Basal Ganglia such as parallel pathways, we find that representing different inputs in different pathways based on the concept of stripes seen in working memory models offer superior accuracy in multi armed bandit tasks over conventional reservoirs.

Abstract

The building sector consumes a significant amount of energy. The three major building performance areas are comfort, energy efficiency, and demand response capabilities. In the recent past, widespread penetration of Distributed Energy Resources (DER) have increased interest in using end-user devices and equipment in buildings as flexible devices to balance grid supply and demand, i.e., �Grid-Responsive� buildings. DER refer small-scale power generation or storage technologies that can be deployed close to the point of energy consumption. These resources are often decentralized and include renewable energy sources such as solar panels, wind turbines, and energy storage systems like batteries. Integrating DER introduces many operational challenges and uncertainty into the grid. Demand Response (DR) strategies are applied to address these challenges. DR refers to the practice of actively adjusting electricity consumption in response to signals such as price or incentives from grid operators, utilities, or aggregators. It involves reducing or shifting electricity usage during peak periods or in response to grid constraints to balance supply and demand, enhance grid reliability, and avoid or mitigate grid emergencies. The approach that is used to manage the DR using price as the key operational parameter is called Transactive Controls (TC). TC is a market-based control paradigm that uses �price� as the key operational parameter. GridWise Architecture Council defined it as �a set of economic and control mechanisms that allow the dynamic balance of supply and demand across the entire electrical infrastructure using value as a key operational parameter�. Economists extensively dealt with TC in microeconomics. However, TC is a domain-free approach that integrates market-based coordination and value-based control for a group of resources to achieve global objectives. Furthermore, the Building Energy Management Systems (BEMS) manage TC at the building or zone level in a commercial building. For example, in case of load shedding, it switches off low priority zone air conditioning or raises the set-point for cooling irrespective of usage in the whole building; dims all the lights regardless of the occupants and their needs. The occupant does not get an opportunity at the time of the DR event to decide on which comfort parameters he is willing to forgo to meet the energy demand. Besides, building energy managers usually operate the buildings to maintain homogeneous indoor ambient conditions (like zone temperatures and lighting). However, occupants have individual thermal and visual comfort preferences. Maintaining a homogeneous indoor environment throughout the building/zones leads to unnecessary energy consumption as well as the inability to meet the comfort needs of the occupants. This has led the building science research community to pursue Personal Environment Control Systems (PECS), such as local thermal conditioning systems like heated computer keyboard, personal heaters, desk fans, and radiant cooling cubicles and task lighting systems such as desk lamps. PECS create favourable micro-ambient conditions around each occupant. Nevertheless, the literature study shows that there is a gap and a need for a system/framework to integrate PECS within the task environment and between task & ambient controls to address the challenges. Hence, as part of this thesis, a system, iSPACE - intelligent System for Personal Ambient Control and energy Efficiency, has been developed to address the challenges. iSPACE integrates transactive control methods to facilitate demand response strategies within personal environmental control systems. This thesis introduces an innovative approach to integrating Personal Environmental Control Systems (PECS) into building energy management, emphasising demand response through transactive controls. It addresses the limitations of traditional demand response mechanisms by empowering individual occupants to manage energy usage at a granular level. The proposed framework establishes a hierarchical distributed multi-agent system architecture, seamlessly integrating PECS devices like SmartHub, RadiantCubicle, and SmartStrip into building energy management systems. The focal point of this thesis is the implementation of transactive control within the building environment, facilitating interactions between energy-consuming devices and the building systems at a localized level. While traditional supply-demand balancing occurs at the grid level by utilities or balancing authorities, our research explores the efficacy of implementing transactive control within buildings to optimize energy usage and enables end-users to participate in managing DR events at the task level based on their priorities. Personalised demand response strategies are facilitated by enabling real-time interaction between occupants and energy-consuming devices. Key contributions include developing and implementing

Abstract

In the literature several parsing systems for natural language processing exist which use different methods to parse sentences. Most of the previous work done in this domain was either done using hand written grammar or by using data-driven methods. Both these methods have pros as well as cons associated with them. As data-driven parsers rely on annotated treebanks they perform at high accuracies when they encounter familiar data that they have been trained on. The only downside to this method is that when they encounter unknown data in a sentence their performance drops drastically. On the other hand, when we come to parsers which use hand written grammars, they have a broad coverage on the language hence we could say that there is no such a thing as seen or unseen data. The only problem with this method is that it is very hard to write a grammar that has broad coverage in a language. In this thesis we develop a method where the con of one method is overcome by the pro of the other. In short we could say that we are building a parser which uses grammar rules extracted from a treebank (which would cover the aspect of broad coverage) and also has the data-driven approach included in it. With this thesis we try to achieve two goals: � Developing a novel approach to parsing text. � Being able to use the resulting parser using very minimal resources.

Abstract

This thesis focuses on the residential electricity demand in India, where the residential sector makes up 27% of the overall electricity usage. The study aims to provide a detailed understanding of the consumption patterns of household appliances. To achieve this, the electricity consumption of individual household appliances in 21 homes across different locations in Hyderabad was monitored. Device-level monitoring equipment, such as smart meters and sockets, was used to collect data for 61 days in summer (1 May 2023 to 30 June 2023) and 30 days in winter (1 Nov 2023 to 30 Nov 2023). The dwellers were chosen on a voluntary basis and had the option to opt out of the study at any point. The study focuses on two aspects of residential electricity consumption patterns: household and appliance levels. For the household-level study, five homes with a greater number of monitored appliances were selected and analyzed for load profiles of regularly used major appliances and daily electricity breakdown during the peak summer (15 May 2023 to 14 June 2023). Moreover, consumption patterns of ACs and geysers were studied individually in detail for the summer and winter. On the other hand, appliance-level analysis was conducted on the appliances, namely washing machine, refrigerator, television and water purifier for summer and winter. Load profiles of these appliances were created to understand seasonality in their usage patterns. The household-level analysis of five homes during peak summer found that six major appliances, namely AC, geyser (domestic water heater), refrigerator, washing machine, television and water purifier, consumed approximately 71% of the total daily electricity. Based on the analysis, it was found that the primary energy consumers in a typical home are the ACs and geysers. During periods of low solar availability, air conditioners accounted for the majority of electricity consumption, while geysers and washing machines were used more frequently during peak solar hours. By implementing a range of strategies, such as adjusting thermostat settings, utilizing fans alongside air conditioners, and opting for high-efficiency models of air conditioners, there is a potential to reduce daily electricity consumption by approximately 10% (437 Wh per home). Another potential measure to consider is the introduction of a district cooling system that incorporates thermal storage, which would shift the energy usage of air conditioners to daytime hours, thus further decreasing nighttime demand. Shifting the geyser usage can redirect 1.05 kWh of daily electrical energy consumption per household during peak solar hours. Additionally, replacing geysers with heat pumps can significantly reduce domestic hot water energy consumption to one-third of the current daily usage. By analyzing these consumption patterns, the study identifies peak consumption periods and explores potential strategies to shift electricity consumption to peak solar hours so that energy from solar PV can be utilized. Moreover, the study also focuses on appliance-level consumption patterns, highlighting the AC and geyser usage trends. It was observed that when the daily average temperatures are greater than 29�C, the energy consumption of ACs tends to be high. Conversely, geysers are utilized in both summer and winter seasons, with a higher energy consumption pattern evident during the winter months. The average load the ACs and geysers imparted during usage was 0.6 kW and 1.87 kW per appliance, respectively. The thesis concludes by providing recommendations on how households can reduce their energy usage and shift their usage patterns to peak solar hours. These include adjusting the setpoints of air conditioners to higher temperatures, using ceiling fans with air conditioners, and utilizing appliances with higher energy efficiency ratios. Ongoing research and adaptation will be crucial to maximize the benefits of these innovations for both consumers and the broader energy ecosystem. The findings of this study provide insights for policymakers, utility providers, and residents looking to reduce energy consumption and achieve a more sustainable energy system. Overall, this thesis report provides a detailed analysis of the electricity consumption patterns of households in Hyderabad and offers practical recommendations for shifting usage patterns to peak solar hours.

Abstract

Text classification and generation are linchpins in the explosive growth of the internet, seamlessly organizing and generating vast volumes of textual information. These pivotal techniques not only enhance Natural Language Processing but contribute indispensably to the dynamic evolution of online communication and knowledge dissemination in the modern world. This is why we have selected Sexism Detection and automated Wikipedia article generation in Indian languages as our topics for researching techniques in text classification and text generation, respectively. Detecting sexism in text promotes inclusivity and gender equality online, whereas generating Wikipedia articles in Indian languages enhances the accessibility of knowledge to diverse communities. With the rapid growth of online communication on social media platforms, there has also been an increase in the amount of hate speech, especially in terms of sexist language online. The proliferation of such sexist language has an impact on the mental health and well-being of the users and, hence, underscores the need for automated systems to detect and flag such pieces of text online. In this thesis, we explore the effectiveness of both conventional machine-learning techniques and different modern techniques applied to BERT-based classifiers for detecting sexist text in the EDOS shared task. The results show that the performance of the baseline conventional classifiers and BERT-based classifiers is greatly improved when techniques like Data Resampling, Data Augmentation, Domain Adaptive Pre-training, and Learning Rate Scheduling are implemented. The surge of data on the internet has also seen the rise of a knowledge giant, Wikipedia. In today�s world, it is seen as one of the primary online sources for the distribution and consumption of free knowledge. Wikipedia users collaborate in a structured and organized manner to publish and update articles on numerous topics. English Wikipedia ranks at the top when it comes to the amount of information available for any language (> 6.7 million articles). However, the amount of information available in Indian languages is lagging far behind. The same article in Hindi may be vastly different from its English version and generally contains less information. This poses a problem for native Indian language speakers who are not proficient in English. Therefore, having the same amount of information in Indian languages will help promote knowledge among such speakers. Publishing the articles manually, which has been the status quo for decades, is a time-consuming process. To get the amount of information in native Indian languages up-to-speed with the amount of information in English, automating the whole article generation process is the best option. In this thesis, we present a stage-wise approach ranging from Data Collection to Summarization and Translation, and finally culminating with Template Creation. This approach ensures the efficient generation of a large amount of content in Hindi Wikipedia in less time. All in all, this work aims to develop automated systems to foster a safe digital space for users of social media platforms and propose a stage-wise approach for the efficient generation of Wiki articles, hence promoting the dissemination of knowledge for native Indian language speakers.

Abstract

Neural Radiance Fields (NeRFs) have emerged as a pivotal advancement in computer graphics and vision. They provide a framework for rendering highly detailed novel view images from sparse multi- view input data. NeRFs use a continuous function to represent scenes that can be estimated using neural networks. This approach enables the generation of photorealistic images for static scenes. Outside the domain of image synthesis, NeRFs have been widely adopted as a representation of several downstream including but not limited to scene understanding, augemented reality, scene nav- igation, segmentation, and 3D asset generation. In this thesis, we explore upon the segmentation and editing capabilities in radiance fields. We propose a fast style transfer method that leverages multi-view consistent generation of stylized priors to change the appearance vectors in a Tensorial Radiance Field. Our method promises a speed-up of several orders of magnitude in applying style transfer and adheres to the colorscheme from the style image better than previous works. Next, we tackle the task of segmentation in radiance fields. Our method uses a grid-based feature field which allows extremely fast feature querying and searching. Combined with our stroke-based seg- mentation, this allws the user to interactively segment objects in a captured radiance field. We improve the state-of-the-art in terms of segmentation quality by a huge margin and in terms of segmentation time by orders of magnitude. Our method enables basic editing capabilities like translation, appearance editing, removal, and composition for which we show preliminary results. We further explore the problem of composition of radiance fields. Composition of two radiance fields using ray marching requires twice the amount of memory and compute. We use distillation to fuse multiple radiance fields into one to circumvent this problem. Our distillation process is roughly thrice as fast as re-training and produces a unified representation for radiance fields.

Abstract

We introduce a novel approach for multimodal story summarization, aimed at leveraging TV episode recaps to create concise summaries of complex storylines. These recaps, which consist of short video sequences combining key visual moments and dialogues from previous episodes, serve as a valuable source of weak supervision for labeling the summarization task. To facilitate this approach, we introduce the PlotSnap dataset, which focuses on two crime thriller TV shows. Each episode in this dataset is over 40 minutes long and is accompanied by rich recaps. These recaps are mapped to corresponding sub-stories, providing labels for the story summarization task. Our proposed model, TaleSumm, operates hierarchically. (i) First, it processes entire episodes by generating compact representations of shots and dialogues. (ii) Then, it predicts the importance scores for each video shot and dialog utterance, taking into account interactions between local story groups. Unlike traditional summarization tasks, our method extracts multiple plot points from long-form videos. We conducted a comprehensive evaluation of our approach, including assessing its performance in crossseries generalization. TaleSumm demonstrates promising results, not only on the video summarization benchmarks but also in effectively summarizing the intricate storylines of the TV shows in the PlotSnap dataset. Our project implementation as well as dataset features and demo can be found at https: //github.com/katha-ai/RecapStorySumm-CVPR2024.

Abstract

In this thesis, we delve into the analysis of movie narratives, with a specific focus on understanding the emotions and mental states of characters within a scene. Our approach involves predicting a diverse range of emotions for individual movie scenes and each character within those scenes. To achieve this, we introduce EmoTx, a novel multimodal Transformer-based architecture that integrates video data, multiple characters, and dialogues for making comprehensive predictions. Leveraging annotations from the MovieGraphs dataset, our model is tailored to predict both classic emotions (e.g., happiness, anger) and nuanced mental states (e.g., honesty, helpfulness). Our experiments concentrate on evaluating performance across the ten most common and twenty-five most common emotional labels, along with a mapping that clusters 181 labels into 26 categories. Through systematic ablation studies and a comparative analysis against established emotion recognition methods, we demonstrate the effectiveness of EmoTx in capturing the intricacies of emotional and mental states in movie contexts. Additionally, our investigation into EmoTx’s attention mechanisms provides valuable insights. We observe that when characters express strong emotions, EmoTx focuses on character-related elements, while for other mental states, it relies more on video and dialogue cues. This nuanced understanding enhances the interpretability and contextual relevance of EmoTx in the domain of movie story analysis. The findings presented in this thesis contribute to advancing our comprehension of character emotions and mental states in cinematic narratives.

Abstract

Blockchain is a distributed ledger and it validates the transactions without any intervention of a trusted third party (TTP). There are several advantages of using the blockchain-based smart grid infrastructure, because decentralization, immutability, transparency, confidentiality and trust are maintained. A blockchain-based smart grid system contains several entities, such as trusted registration authority (RA), service providers (SPs), IoT-enabled smart meters (SMs), and users associated with a smart meter. SPs organize the electricity allocation and energy trading system, and SMs are responsible for monitoring the power utilization and they maintain the pricing to the consumers (users). SMs can be deployed in the homes, and an attacker may capture some SMs and use the secure data stored into it. The communications between SPs and SMs must be secure so that passive/active attacks should not be possible. To ensure the security and privacy of the users’ private information, it is extremely required to design a secure and efficient access control scheme between SMs and SPs. With the help from the blockchain technology, the secure data can be stored in the form of blocks in a private blockchain. The SPs involved in the P2P SP network are responsible in validating the new blocks before adding them into the blockchain using the consensus algorithm. To mitigate these issues, we first propose a new blockchain-based access control protocol in internet of Things (IoT)-enabled smart-grid system, called DBACP-IoTSG. Through the proposed DBACP-IoTSG, the data is securely brought to the service providers from their respectively smart meters. The Peer-to-Peer (P2P) network is formed by the participating services providers, where the peer nodes are responsible for creating the blocks from the gathered data securely coming from their corresponding smart meters and adding them into the blockchain after validation of the blocks using the voting-based consensus algorithm. In our work, the blockchain is considered as private because the data collected from the consumers of the smart meters are private and confidential. By the formal security analysis under the random oracle model, non-mathematical security analysis and software-based formal security verification, DBACP-IoTSG is shown to be resistant against various attacks. We carry out the experimental results of various cryptographic primitives that are needed for comparative analysis using the widely-used Multiprecision Integer and Rational Arithmetic Cryptographic Library (MIRACL). A detailed comparative study reveals that DBACP-IoTSG supports more functionality features and provides better security apart from its low communication and computation costs as compared to recently proposed relevant schemes. In addition, the blockchain implementation of DBACP-IoTSG has been performed to measure computational time needed for the varied number of blocks addition and transactions per block in the blockchain.The Industrial Internet of Things (IIoT) is able to connect machines, analytics and people with IoT smart devices, gateway nodes and edge devices to create powerful intuitivenesses to drive smarter, faster and effective business agreements. IIoT having interconnected machines along with devices can monitor, gather, exchange, and analyze information. Since the communication among the entities in IIoT environment takes place insecurely (for instance, wireless communications and Internet), an intruder can easily tamper with the data. Moreover, physical theft of IoT smart devices provides an intruder to mount impersonation and other attacks. To handle such critical issues, we next design a new private blockchain-envisioned access control scheme for Pervasive Edge Computing (PEC) in IIoT environment, called PBACSPECIIoT. We consider the private blockchain consisting of the transactions and registration credentials of the entities related to IIoT, because the information is strictly confidential and private. The security of PBACS-PECIIoT is significantly improved due to usage of blockchain as immutability, transparency and decentralization along with protection of various potential attacks. A meticulous comparative analysis exhibits that PBACS-PECIIoT achieves greater security and more functionality features, and requires low costs for communication and computational as compared to other pertinent schemes. The communication among various entities in an edge computing based generic IoT environment takes place via insecure (public) channel (e.g., via the Internet). It gives an opportunity to an adversary to mount various types of attacks, including “replay”, “man-in-themiddle”, “impersonation”, and “Ephemeral Secret Leakage (ESL)” attacks. Moreover, the adversary can physically capture to some IoT smart devices in order to compromise secure communication among other non-compromised nodes in the network. Therefore, it is very much essential that the data n

Abstract

In the dynamic landscape of the semiconductor field, the application and significance of ultra-low power circuits have become paramount, particularly in the domains of the Internet of Things (IoT) and biomedical applications. The relentless pursuit of higher speeds in traditional semiconductor designs is giving way to a new paradigm where energy efficiency takes precedence. In IoT applications, where devices are often constrained by the absence of continuous power sources, ultra-low power circuits enable prolonged operation through energy harvesting or miniaturized batteries. This not only extends the device lifespan but also aligns with the energy-conscious demands of IoT ecosystems. In the realm of biomedical applications, the importance of ultra-low power circuits is underscored by the need for minimally invasive, long-term implantable devices. These circuits facilitate precise sensing, control, and communication within the human body, ensuring both longevity and safety. As the semiconductor industry navigates this transformative shift, the development and integration of ultra-low power circuits emerge as a crucial frontier, shaping the future of semiconductor technology with profound implications for IoT and biomedical advancements. In ultra-low power IoT applications, as well as biomedical implants, need low power consumption, low supply voltage, and high accuracy. Keeping all these constraints in mind, we worked on designing an ultra-low power Resistance-to-Digital Converter (RDC) with improved Figure-of-Merit (FoM) and low-temperature sensitivity for miniature low-powered sensing system. The proposed RDC system consumes only 162nW, with FoM as low as 0.845pJ/conversion cycle. The RDC can measure wide range of resistances measuring from 50k? to 1M?. Due to this, the RDC can be configured to use as different sensors like pressure sensors, temperature sensors, touch sensors, etc. For almost all IoT and biomedical systems, biasing circuits are essential, as these are in general always-on circuits. It makes sense to design voltage and current references in the ultra-low power domain. Consequently, we focused on designing a 2.3nW, sub-Bandgap voltage reference. It generates a reference voltage of 336mV without incorporating any resistors and operates for a high-temperature range of -40?C to 150?C and a supply range of 0.7V-4V. In the above temperature and supply ranges, the proposed circuit�s power consumption stays fairly linear, eliminating the exponential power consumption issue at such high temperatures. Designed in 65nm CMOS process, it achieves an impressive line sensitivity of 0.0066%/V for a supply range of 0.7V to 4V and a PSRR of 89dB at DC and 1V supply. As we all know, the voltage reference always preferred to be process independent, and the proposed circuit shows a �3?-inaccuracy of 4.295% without additional trimming circuits. It used gate-leakage-based resistance to save significant silicon area of high-value resistance. However, we need to design voltage and current references separately, and low-level current references require high-value resistances, and other CMOS current references suffer from high process variation. Consequently, we focused on designing ultra-low power trim free voltage/current reference without using resistance and amplifiers. The whole system works for as low as 37nW. The proposed voltage/current reference outputs a voltage and current of 820mV and 3.23nA, respectively. Without trimming, the process variation of the proposed voltage/current reference is 1.34%(?/�) / 1.75%(?/�). Designed in 90nm CMOS process, it achieved proper trim-free, low-power operation without using resistors, Which reduces area significantly. These attributes make it a desirable choice for various ultra-low power wearables and IoT applications.