Abstract

In this thesis, we address the age old problem of dealing with scenarios in which the significance of making a mistake while taking a classification decision is very high. We also look deeper into how we can take advantage of deep learning techniques to establish state of the art results and study the representations learnt by the deep reject option classifiers. While supervised techniques have their advantage we explored unsupervised learning in real time scenario of catching malicious traders in Stock Exchanges which are establishments where significance of mistakes are very high. Firstly, we propose deep architectures for learning binary instance specific abstain (reject option) classifiers. To show the effectiveness of the proposed approach, we experiment with several real world datasets. We observe that the proposed approach not only performs comparable to the state-of-the-art approaches, it is also robust against label noise. We also provide visualizations to observe the important features learned by the network corresponding to the abstaining decision. Secondly, we look at Stock exchanges. Due to their prominence, stock exchanges are prone to a variety of attacks. Different types of proprietary fraudulent activity detectors are deployed by stock exchanges to analyze the time series data of trader’s activities or the activity of a particular stock to flag potentially malicious transactions while human analysts probe the flagged transactions further. The key issue faced here is that while the number of anomalous transactions identified can run into thousands or tens of thousands, the number of such transactions that can realistically be probed by human analysts would be a small fraction due to resource constraints. The issue therefore reduces to a dynamic resource allocation problem wherein alerts that represent the most malicious transactions need to be mapped to human analysts for further probing across different time intervals. Thirdly, we develop a multiclass deep learning solution with a rejection option. While binary so- lutions find relevance in scenarios such as cancer detection where ”yes” or ”no” could be the probable options. In scenarios such as autonomous driving, multiple options could be applicable at each instant. So, we propose a multiclass double ramp loss function. We study the properties of the loss function and establish state of the art results on real world datasets.

Abstract

Binocular vision (BV) is the result of the fusion of the input from each eye to form a coherent image. BV is essential for an individual to have a normal vision. Clinically BV anomalies (strabismus, amblyopia) are evaluated using different diagnostic tests and instruments. In recent studies, non-strabismus BV anomalies have also been reported and flagged to be included in conventional screening protocols. A traditional instrument is the Synoptophore which evaluates three grades of BV. This equipment is very efficient but has certain limitations a) ambient light in the testing room, b) bulky and hence difficult to transport to screening camps in remote areas, c) Cost of Synoptophore is another major disadvantage, making it out of reach of most eye care centers etc. The Aim of this study is to address these limitations with existing technology and provide a suitable solution for the same. Virtual Reality(VR) is one such rapidly evolving technology which provide a suitable platform to implement the principle of Synoptophore easily. Therefore, we propose VR-Phore, a VR application solely based on principle of the haploscope as in the Synoptophore. Which uses a head-mounted display device (Oculus Rift is used for this study) for diagnostics and also as possible therapeutic use. The proposed system addresses the limitations of the Synoptophore. Limitations such as requirement of suiatable light source in the room while using synoptophore, there is no such requirement for VR-Phore usage. VR-Phore is a simple VR application which can be easily set up at any places like screening camps in schools, clinics or in any remote areas etc. Operating Synoptophore requires a clinical professional with prior knowledge on the device whereas VR-Phore can be used by anyone and the stored data can be provided to doctors for diagnosis. VR-Phore also provide added advantages of a software platform to incorporate testing modules for a range of clinical conditions very easily and tele-diagnosis. A total of 41 participants are involved in this study for data collection. The data collected from this system was compared to Synoptophore. Measurements, high correlations and accuracy was observed validating the application of the proposed system in diagnostics. The added advantage is the portability of the software platform onto a smart phone attached to a low-cost VR head-mounted device, effectively making this system cost-friendly and affordable to everyone.

Abstract

Online abuse and offensive language on social media have become widespread problems in today’s digital age. With the growing use of the web in our daily lives, identifying and countering such speech has become a popular topic in the research community. Authorities of many countries worldwide recognise this as a serious problem, particularly because the reach of the internet has grown exponentially over the past years. Most previous work in this area has focused on detecting such harmful speech to prevent users from experiencing the adverse impact of the same. Detection is achieved chiefly through text classification models, which learn to recognise hate speech/offensive speech from neutral/standard sentences. More recently, with the advent of style transfer in text, many such negative attributes of speech are transferred into positive ones. One such instance is sentiment transfer, in which negative reviews are rewritten as positive ones while attempting to preserve the non-stylistic content of the input. In this thesis, we contribute the JDC (Jibe and Delight Corpus), a Reddit-based dataset consisting of 68, 159 insults and 51, 102 compliments. Our dataset is unique and differs from typical online abuse datasets since instead of being targeted at a community or race (in case of hate speech), they are targeted at individuals, and exhibit “creativity” in their formation, and so do not rely on profanity or slurs to convey the negative intent of the insult. We also perform extensive experiments on the same on two different fronts. We first work on the detection of such speech by analysing the results of various types of text classification models. Secondly, we propose to convert such insults into compliments by using unsupervised style transfer. We benchmark multiple existing state-of-the-art models for both text classification and unsupervised style transfer on the dataset. Our best classification model shows an accuracy of 97.8% and an F1 score of 0.972. We analyse the experimental results and conclude that the transfer task is challenging, requiring the models to understand the high degree of creativity exhibited in the data.

Abstract

With extensive use of the Internet in recent years, massive amounts of data related to different topics and developments have been generated and shared online. The advent of social media like Twitter, Facebook, and Reddit has accelerated the communication between people of all colors, nations, and languages. Though the platform exists, communication barriers still exist due to languages. Many researchers are trying to solve this through various methods. India is a land of multiple languages, and a majority of people are multilingual and tend to mix words from different languages in written text and also in speech. India has twenty-three significant languages with over seven hundred and twenty dialects. Kannada is a Dravidian language, and it is one of the four main Dravidian languages, and it is one of the top 30 most spoken languages of the world, with its own independent script and over fifty million speakers. With recent surges in developments in machine learning, understanding the language, and thus resulting in a wide variety of applications for resource-rich languages. In resource-rich languages, a considerable amount of work has been performed. The same is not true for low-resource languages or even Code-mixed low-resource languages. Kannada is one of such low-resource languages. In this thesis, we aim to focus on the research areas related to the Kannada-English Code-mixed domain. We present corpora in Kannada-English Code-mixed Twitter data and their analysis in the areas of Emotion Prediction and Parts-of-Speech Tagging, respectively. In the first part, we present the corpus for Emotion Prediction in Kannada-English Code-mixed data. This is, to the best of our knowledge, the first corpus in this field. Identification and analysis of emotions in user-generated data in social media like Twitter, Facebook, Reddit, etc., is essential in understanding the daily trends and human behavior. There has been work done on Code-mixed social media corpus but not on emotion prediction of Kannada-English Code-mixed Twitter data. We analyze the problem of emotion prediction on corpus obtained from Code-mixed KannadaEnglish extracted from Twitter annotated with their respective ‘Emotion’ for each tweet. We experimented with machine learning prediction models using features like Character N-Grams, Word NGrams, Repetitive characters, and others on Support Vector Machines(SVM) and Long Short term Memory Network(LSTM) on our corpus. In the second part, we present the corpus in Kannada-English Code-mixed data for Parts-of-Speech Tagging. This is, to the best of our knowledge, the first corpus in this field. Part-of-Speech (POS) is essential for many Natural Language Processing (NLP) applications. There has been a significant amount of work done in POS tagging for resource-rich languages. POS tagging is an essential phase of text analysis in understanding the semantics and context of language. These tags are useful for higher-level tasks such as building parse trees, which can be used for Named Entity Recognition, Coreference resolution, Sentiment Analysis, and Question Answering. There has been work done on Code-mixed social media corpus but not on POS tagging of KannadaEnglish Code-mixed data. Here, we present Kannada-English Code-mixed social media corpus annotated with corresponding POS tags. We also experimented with machine learning classification models Conditional Random Fields(CRF), Bidirectional Long Short term Memory Network(Bi-LSTM), and Bidirectional Long Short term Memory Network-Conditional Random Fields(Bi-LSTM-CRF) models on our corpus.

Abstract

Energy consumption of the residential sector is increasing year after year with economic development, urbanization, and improvement of people’s living standards. In order to achieve more efficient energy consumption, it is crucial to provide appliance-wise energy consumption feedback to the users to save energy. Appliance wise energy consumption feedback allows users to identify faulty appliances, reduce energy demand and identify unnecessary active devices. Individual appliance energy consumption is monitored by two approaches intrusive load monitoring and non-intrusive load monitoring. Intrusive load monitoring uses smart plugs and smart strips on each device to monitor their power consumption.Non-intrusive load monitoring (NILM) estimates the individual appliance power consumption from the main meter aggregated data without using additional sensors. NILM is a cost-effective approach for giving appliance-wise energy consumption feedback. Non-intrusive load monitoring (NILM) is a blind source separation problem that requires a system to estimate the electricity usage of individual appliances from the aggregated household energy consumption. In this thesis, we will discuss about NILMmethods and algorithms on low sampled (1 min) aggregated data in detail and identify the research gaps. To addresses those research gaps, we propose a novel deep neural network-based approach for performing load disaggregation on low-frequency power data (1 min) obtained from residential households. We combine a series of one-dimensional Convolutional Neural Networks and Long Short Term Memory(1D CNN-LSTM) to extract features that can identify active appliances and retrieve their power consumption given the aggregated household power value. We took appliances which are multi-state appliances(washing machine, dishwasher, microwave, and refrigerator ) to test the algorithm. The disaggregation performance of the algorithm is measured using six metrics and compared with five stateof-the-art algorithms. To explore how well the algorithms generalize to unseen houses, the performance of the algorithms was measured in two separate scenarios: one using test data from a house not seen during training and a second scenario using test data from houses that were seen during training. Our neural net achieves better F1 scores (across all four appliances) than state-of-the-art algorithms and generalizes well to unseen houses with a lower number of trainable parameters. The algorithm is designed for low-power offline devices. Empirical calculations show that our model outperforms the state-of-theart on Reference Energy Disaggregation Dataset (REDD) and UK-dale dataset. We have collected the electricity consumption data of 11 houses in India for 19 days as part of the Residential Building Energy Demand in India (RESIDE) Project. The proposed model on RESIDE dataset achieved better F1 scores and generalised well on the unseen houses.

Abstract

In communication systems, the reliable transmission of message bits is an important factor in the overall efficiency of the system. The correction of errors encountered in transmitted messages is done with the use of error correcting codes. These codes have traditionally been designed with concepts of abstract algebra. In the coding theory domain, the pursuit of good decoders and channel code designs has always been an interesting area of research. Recently because of the use of advanced GPU machines, neural networks have become a tool of choice for modern day researchers to do various tasks like function approximations and algorithm learning. There has been a surge in the research work trying to develop end to end codes and decoders for channel coding applications. Explorations into the modulation and demodulation designs for various channel settings were the first to come up in the applications of neural networks for channel coding. Following this came up the design of decoders for various kinds of existing codes in the coding theory. Then, explorations into the design of end to end communication systems and autoencoders came up in the literature. We have come up with a design of a decoder for a coded modulation technique called as TCM (trellis coded modulation). The design is based on RNN and CNN variants of neural networks and shows that neural networks can perform the task of decoding coded modulation messages as effectively as the existing algorithms for codeword decoding. The trained model is tested for robustness to channel models not seen while training, and the proposed decoder is better in performance as compared to the existing decoder. An interpretation of the working of the proposed decoder has also been presented in this thesis, where we show the classification of symbols into decision regions with the use of CNN along with softmax activation function. In another work, we have implemented an autoencoder based on neural networks for product code topology. The autoencoder takes ideas from turbo product codes for its design. The idea behind its operation is that the code is divided into its component row and column codes, the operation of the encoding is done by applying row and column codes one after another to generate the codeword. The decoder utilises the extrinsic information extraction equations to do the decoding of the received codeword. The received codeword is decoded in a row wise and column wise manner to get confidence values for every bit in the transmitted codeword; these are then used in the information extraction equations. The training of the autoencoder is done for the AWGN channel and it shows performance improvement over uncoded BPSK modulated messages. The encoder performance can be improved with the use of higher codeword lengths while training. With the use of individually trained row and column component codes in the autoencoder, the performance of the proposed autoencoder may get further improved.

Abstract

Biological neurons are prototypical examples of nonlinear excitable systems. Nonlinear dynamics and a strong coupling of the relevant variables allows neurons to function in an array of different physiological modes. Such complexity at the single neuron level is ultimately responsible for the immense complexity of the mammalian nervous system. From a mathematical standpoint as well, neuron models are highly of interest as they provide opportunities to explore the larger domain of excitable systems and threshold phenomena. As a result, theoretical and computational studies of neuron models have in the last few decades, explained and predicted interesting and unexpected phenomena, as seen for instance when external noise interacts with the dynamics of an excitable system. The most prominent aspect of neuronal information processing is the phenomenon of synaptic integration. The detailed mechanisms that modulate synaptic inputs determine the computational properties of any given neuron. In this thesis, we study a simple model for the summation of excitatory inputs from synapses and illustrate its use by characterizing some functional properties of postsynaptic neurons. Our experiments are based on the well studied FitzHugh-Nagumo (FHN) model for action potential generation. The FHN equations have been investigated extensively since their introduction in the early 1960s. In particular, a number of insights into the dynamics of excitation thresholds have been gained from this model as well how these thresholds change as bifurcations are approached. The FHN model is also well known for its ability to exhibit interesting noise driven properties like stochastic resonance. By applying minimal perturbations to the FHN system, we provide a simple framework to study the dynamics of postsynaptic neurons. The model is based on a first order approximation of dendritic currents as identical rectangular current pulses. Using this, we obtain firing threshold curves of postsynaptic neurons as a function of the average time gap between successive dendritic currrent pulses. Furthermore, we explain the features of these threshold curves using an analysis based on nullclines and a class of trajectories known as Canards. A realistic simulation of an individual neuron involves different spatial and temporal scales resulting in computationally intensive/prohibitive requirements even for a modestly small number of dendrites especially if we want to investigate long-time statistical properties of the firings. This is what motivates us to introduce a minimal model for synaptic integration. Although not biophysically realistic on finer scales, we try to show that a simple model can nonetheless highlight the general statistical features of spike timings in neural processes where the active properties of dendrites do not play a major role. In particular, our model postsynaptic neurons are especially suited to study how a simple integration mechanism such as the spatial summation of dendritic currents can lead to nontrivial dynamics. Noise is a ubiquitous presence in neuronal processes and has been shown to play important functional roles in sensory physiology among other things. To study the functional aspects of noise in sensory information processing, we obtain the firing statistics of postsynaptic neurons when their corresponding presynaptic neurons are subjected to two well known noise driven processes: stochastic and coherence resonance. These phenomena have been observed in a variety of models as well as experimental systems. By subjecting the model to the conditions of these processes, we propose some advantages of postsynaptic neurons towards processing weak or irregular sensory stimuli. Interspike interval histograms are employed to demonstrate the contrast between presynaptic and postsynaptic neurons in biological scenarios such as the transmission of sensory information. Biologically relevant implications of the results are highlighted throughout and further experiments suggested. The model requires a small number of parameters and is especially useful to isolate the role of integration mechanisms that rely on summation of inputs with little dendritic processing. Simple extensions/modifications to include more aspects of synaptic integration such as inhibitory synapses, bursting dynamics, etc are also dealt with. The simplicity of the model establishes a base case in a bottom-up modelling hierarchy while maintaining the ability to explore new functional roles of postsynaptic neurons

Abstract

DNNs (Deep Neural Networks) have found use in a variety of applications recently, and have become much larger and resource-hungry over the years. CNNs (Convolutional Neural Networks) are widely used in Computer Vision tasks, which make use of the convolution operation in successive layers – making them computationally expensive to run. Due to the large size and computation required to run modern models, it is difficult to use them in resource-constrained scenarios, such as mobiles and embedded devices. Also, the amount of data created and collected is growing at a rapid rate, and annotating large amounts of data to use it for training DNNs is expensive. There are approaches which seek to intelligently query samples to train upon in an iterative fashion, such as Active Learning – but AL setups are themselves very costly to run due to full training of large models many times in the process. In this thesis, we explore methods to achieve extremely high speedups in both CNN model inference, and train-times for AL setups. Various paradigms to achieve fast CNN inference have been explored in the past, two major ones being binarization and pruning. Binarization involves the quantization of weights and/or inputs of the network from 32-bit full precision floats into a {-1,+1} space, with the aim of both achieving compression (as singular bits occupy 32 times less space than 32-bit floats) and speedups (as bit-bit operations can be done faster). Network pruning, on the other hand, tries to identify and remove redundant parts of the network in an unstructured (individual weights) or structured (channels/layers) manner to create sparse and efficient networks. While both these paradigms – binarization and pruning – have demonstrated great efficacy in achieving speedups and compression in CNNs, little work has been done in attempting to combine both approaches together. We argue that these paradigms are complementary, and can be combined to offer high levels of compression and speedup without any significant accuracy loss. Intuitively, weights/activations closer to zero have higher binarization error making them good candidates for pruning. We propose a novel Structured Ternary-Quantized Network that incorporates speedups from binary convolution algorithms through structured pruning, enabling the removal of pruned parts of the network entirely post-training. Our approach beats previous works attempting the same by a significant margin. Overall, our method brings up to 89x layer-wise compression over the corresponding full-precision networks – achieving only 0.33% loss on CIFAR-10 with ResNet-18 with a 40% PFR (Prune Factor Ratio for filters), and 0.3% on ImageNet with ResNet-18 with a 19% PFR. We also explore the field of AL (Active Learning), which is used in scenarios where unlabelled data is abundant, but annotation costs for that data make it infeasible to utilize all data for supervised training. AL methods initially train a model with a small pool of annotated data, and use the model’s (referred to as the selection model) predictions on the rest of the unlabelled data to form a query for annotating more samples. The training-query process iteratively happens till a sufficient amount of data is labelled. However, in practice, this sample selection practice in AL setups takes a lot of time as the selection model is fully re-trained on the labelled pool every iteration. We offer two improvements to the standard AL setup to bring down the overall train time required for sample selection significantly: the first, the introduction of a “memory” that enables us to train the selection model for fewer samples every round as opposed to the entire labelled dataset so far, and the second, the use of fast convergence techniques to reduce the number of epochs the selection model trains for. Our proposed improvements can work in tandem with previous techniques such as the use of proxy models for selection, and the combined improvements can bring more than 75x speedups in overall sample selection time to standard AL setups, making them feasible and easy to run in real-life scenarios where procuring large amounts of data is easy, but labelling them is difficult.

Abstract

Over the past few decades, there has been substantial growth in electricity demand of urban buildings across the globe due to rapid urbanisation and usage of energy-intensive systems. Buildings account for over 1/3rd of global energy usage and nearly 40% of energy-related carbon emissions. Due to this fact, energy-efficient buildings with appropriate design standards are adopted by the architects. Performing modelling and simulation using software such as EnergyPlus to assess the scope of energy consumption of a building requires expertise. Web-based parametric building energy simulation tools can help architects perform their desired studies without having much software expertise. Traditional web-based tools use a synchronous request-response method. It is utilized for smaller Input/Output CPU interactions between the web client and the web server. It blocks the client until the task assigned is completed by the server. EnergyPlus is CPU-intensive and multiple simulation requests on the web server can become a bottleneck to the synchronous approach. The parallel processing methods have been explored by the researchers to address this issue but the development of scalable and fault-tolerant systems have not been discussed so far. In this regard, two domain orientated web applications have been developed using asynchronous Distributed Task Queues (DTQ) in this work. The first application is a simple roof savings calculator where the benefits of a cool roof, sloped roof and green roof are evaluated over a conventional roof using EnergyPlus software. There are several independent cool roof and green roof calculators available online which are either look-up table or synchronous simulation engine based. They are limited to certain weather data and can block multiple users accessing the application simultaneously. The roof savings tool developed in this work uses DTQs which brings everything at the same place by providing an interface to compare the cool roof, sloped roof and the green roof over a conventional roof on every location provided by the EnergyPlus website which are over two thousand weather locations. It also includes a parametric simulation option which can let users decide the thickness of the roof based on the energy savings potential. This is the only tool available online for world-wide usage having such diverse simulating options and is meant to handle peak calculator usage. In the second application, the DTQ technique is implemented considering the notable complexity of early-stage design optimisation of a building. During the early stages of building design, architects tend to evaluate various strategies to simulate building energy consumption. There can be several numbers of building parameters to be evaluated such as passive roof strategies, window glazing material, the orientation of the building, window overhang depth, window-to-wall ratio, etc. In a situation having five-building parameters with ten design alternatives for each parameter, the web server needs to evaluate a hundred thousand simulation models with building energy simulation software. With each simulation model taking one minute to process, it may take about seventy days on a single-core system to complete the group of simulations. Using asynchronous DTQ technique, computational cores can be scaled based on the simulation requirements. DTQs does real-time processing while supporting task scheduling and CPU resource management as well. Early Design Optimisation tools (eDOT) for Air-Conditioned buildings have already existed in the past but the development of fault-tolerant systems has not been mentioned yet. DTQ software like Celery provides scalable and reliable features. Even if any of the systems go offline, the rest of the systems can still communicate with the Queue and perform tasks of the disconnected systems. There is an increasing concern of energy efficiency in buildings due to high dependency on mechanical systems for cooling and heating. In this work, the eDOT tool is enhanced by incorporating a Mixed-Mode simulation option. This hybrid approach for space conditioning uses natural ventilation and utilizes mechanical systems only when required to reduce energy consumption without compromising on occupant comfort. Energy Management System (EMS) algorithms are developed for Mixed Mode building models in EnergyPlus to control cooling loads, heating loads and natural ventilation based on certain conditions which include zone air relative humidity, outside relative humidity, zone air temperature, outside air temperature and cooling/heating set points. Parametric simulations for annual building energy consumption are carried out with the eDOT tool for New Delhi and San Francisco weather data. Design strategies with lowest energy profiles for AC and Mixed-mode buildings are examined and compared for both the cities. The cut-off Energy per Intensity (EPI) for identifying strategies

Abstract

Autonomous navigation is a critical task for building intelligent embodied agents. For a successful visual navigation algorithm, building spatial representations for efficient exploration and exploiting structural priors is necessary irrespective of the navigation task. Imagine you are in a house which you have never seen before and you are given the task of "Finding a sink" in that house. While there can be multiple directions to move, most of us would choose the path that would take us to the Dining Room, where the sink is highly likely to occur in the Kitchen. This is because as a human we use strong structural priors to navigate in an unseen environment. This work incorporates these semantic priors on the relative arrangement of objects in a scene as part of our scene understanding module for solving the navigation task effectively. In this thesis, we address the highly challenging problem of object goal navigation in which the agent is tasked to reach any instance of the specified goal category within a defined number of timesteps. The agent, in an unseen environment, has to perceive its surroundings to identify and navigate towards potential regions where the specified goal category can occur. Rather than developing goal driven exploration policies, we aim to adapt the existing exploration policies that maximize scene coverage to be goal-conditioned. Thus, we propose a standalone scene understanding module to identify potential regions where the goal occurs. We also propose Goal-Conditioned Exploration (GCExp), an algorithm that entails the integration of our novel scene understanding module with any existing exploration policy. We test our solution in photo-realistic simulation environments using state of-the-art exploration policy, Active Neural Slam [10], and show improved performance over the same on every evaluation metric. Additionally, we show a marginal improvement on the newly proposed evaluation metric, "Average Categorywise Success", to address the shortcomings of the existing metrics.