Computer Sciencehttp://hdl.handle.net/123456789/52024-02-11T14:17:22Z2024-02-11T14:17:22ZOn the information flow in undirected unicast network (PhD)Qureshi, Mohammad Ishtiyaqhttp://hdl.handle.net/123456789/4672022-09-06T22:05:32Z2022-02-01T00:00:00ZOn the information flow in undirected unicast network (PhD)
Qureshi, Mohammad Ishtiyaq
One of the important unsolved problems in information theory is the conjecture that network coding has no rate benefit over routing in undirected unicast networks. If the conjecture is true then the undirected unicast net- work information capacity is the same as the routing capacity. However, the conjecture is unsolved and the undirected unicast network information capacity is not characterized yet. Even upper bounding the symmetric information rate is a challenging problem. Only two explicit upper bounds on symmetric information rate are known for general undirected networks: (1) sparsest cut bound on symmetric rate is a trivial bound on both commodity and information flow and (2) the linear programming bound using Shannon-type inequalities is generally not used for evaluation due to prohibitively large problem size.
In this work, we characterize an upper bound, called the partition bound, on the symmetric rate for information flow in general undirected unicast networks and present a partitioning technique to obtain converse results for undirected network information flow. We give two proof methods for the partition bound. This bound is further generalized for non-symmetric
rates. We show that the partition bound is not tight in general and also demonstrate an approach to tighten the bound. As a result, we present an alternative proof of the undirected unicast network information capacity of the well known Hu’s 3-pairs network. We give explicit routing solutions achieving the partition bound for (1) two classes of complete n-partite networks called Type-I and Type-II n-partite networks, and (2) a class of
3-layer networks called Type-I 3-layer networks. These results prove that the undirected unicast network coding conjecture holds for these classes of networks. A parameter is defined as an optimal partition which delivers the partition bound. We present a procedure to compute a lower bound for this parameter. This lower bound renders a computable upper bound for the partition bound. We also show that the decision version problem
of computing the partition bound is an N P-complete problem. Thus, both the upper bounds, the sparsest cut bound, and partition bound are not polynomial-time computable unless P=N P. Recently, the undirected unicast network coding conjecture was proved for a new class of networks and it was shown that all the network instances for which the conjecture is proved previously, and the cut based bound is not achievable by commodity flow, are elements of this class. The conjecture was also proved for all undirected unicast networks (1) with six or less number of nodes and (2) with up to three sessions and seven nodes except one particular network. We show the existence of a Type-I n-partite network for which the partition bound is tight and achievable by routing and is not an element of this class of networks. This result establishes that there exist networks outside of the class of networks with unverified conjecture such that the partition bound is tight and attainable by routing.
2022-02-01T00:00:00ZClassification of acoustic scenes using background and foreground (MS)D., Dhanunjaya Varmahttp://hdl.handle.net/123456789/4642022-09-06T22:05:15Z2021-11-15T00:00:00ZClassification of acoustic scenes using background and foreground (MS)
D., Dhanunjaya Varma
The objective of acoustic scene classification is to classify environments based on the sound events they produce. Acoustic scene classification has been used in a variety of applications, which include audio surveillance, assistive technologies like hearing aids and context-aware services. ASC is a challenging task due to the presence of similar sound events across acoustic scenes, causing high inter-class similarity. In this thesis, we approach this problem by providing a mechanism that helps in deriving discriminative features by suppressing certain sound events.
An acoustic scene can be viewed as a combination of background sound events
and foreground sound events. Often, either the background or the foreground carries
beneficial information in identifying the acoustic scenes uniquely. We propose to handle these similar sound events by utilizing a combination of methods that include robust principal component analysis (RPCA), subspace projection techniques and a self-attention network. These methods help in separating the background and the foreground sound events, and in partially removing the background (or foreground) sound events.
We employ the framework of RPCA to decompose the given acoustic scene into the background and the foreground sound events. RPCA decomposes a given data matrix into a low-rank and a sparse matrix. In the context of data describing an acoustic scene, the low-rank matrix represents the slow-changing background, and the sparse matrix represents the occasional foreground sound events. Further, we utilize a subspace projection technique named nuisance attribute projection (NAP) to reduce the inter-class similarity. NAP helps in partially removing the background (or the foreground) sound events by treating either the background (or the fore- ground) as nuisance variations. The nuisance basis for applying NAP are learned from the background and foreground separated data obtained post RPCA. These background-suppressed and the foreground-suppressed representations are combined using fusion techniques to improve classification accuracy. We also present an ap-proach to incorporate the label information in the subspace projections by learning class-specific nuisance bases. Further, projection using these bases in combination with an attention mechanism is used for effective suppression, leading to better discrimination. Our results on standard datasets indicate that the proposed methods that use RPCA and subspace projections are indeed helpful in improving the classification accuracy.
2021-11-15T00:00:00ZModelling air quality via machine learning and IoT technologies (MS)Saini, Tusharhttp://hdl.handle.net/123456789/4632022-09-06T22:05:16Z2022-03-01T00:00:00ZModelling air quality via machine learning and IoT technologies (MS)
Saini, Tushar
Air quality is degrading in developing countries, and severe health issues are associated with increasing air pollution. It is imperative to monitor and predict air quality online in real-time. Although offline air-quality monitoring using hand-held devices is common, online air-quality monitoring is still expensive and uncommon, especially in developing countries. Also, existing technologies do not perform forecasting of air pollution ahead of time. This thesis's first objective was to address these literature gaps and propose a scalable, low-cost real-time air quality monitoring, prediction, and warning system (AQMPWS). The proposed AQMPWS monitored and predicted seven pollutants: PM1.0, PM2.5, PM10, carbon monoxide, nitrogen dioxide, ozone, and sulphur dioxide. The AQMPWS also monitored five weather variables: temperature, pressure, relative humidity, wind speed, and wind direction.
This thesis' second objective involved the forecasting of air pollution ahead of time. Individual and ensemble univariate and multivariate time-series forecasting models were developed and tested for predicting air pollution, which could forecast one step ahead in time (i.e., forecasting was done on the time dimension). Five individual time-series forecasting models, namely, multilayer perceptron (MLP), convolution neural network (CNN), long shortterm memory (LSTM), and seasonal autoregressive integrated moving average (SARIMA), were investigated. Also, a new weighted ensemble model of these individual models was developed. Among the individual univariate models, the CNN performed the best, and this model was followed by the LSTM, MLP, and SARIMA models both during training and test. The investigation revealed that multivariate models performed better than their counterpart univariate models. Furthermore, the weighted ensemble model performed the best among all models.
In the third objective, the developed AQMPWS was benchmarked against an industrialgrade air-quality monitoring system deployed at the ACC Cement Factory in Barmana, India. A high correlation between the collected data from AQMPWS and the industrial-grade system showed that the developed system was on par with the industrial system. Three statistical models, namely, Vector Autoregressive (VAR), Vector Autoregressive Moving Average (VARMA), and SARIMA and an ensemble model, were developed to predict the future values of PM2.5 and PM10 in the industrial system from the PM2.5 and PM10 values measured by the AQMPWS. Results showed that the ensemble model performed the best, and VAR performed the second-best in predicting the PM2.5 and PM10 values in the industrial system.
Lastly, we performed a preliminary study to find the public perception of the adverse effects of air pollution. About 90 percent of those surveyed (N = 500) requested a daily update on their localities' air pollution. Results showed that people preferred periodic SMS alerts to learn about air quality compared to other methods. Results from the study helped shape the alerting system in the AQMPWS.
2022-03-01T00:00:00ZMedical imaging techniques for transformation and inference using deep learning(MS)Srinivasan, Preethihttp://hdl.handle.net/123456789/4612022-08-01T22:08:10Z2021-02-01T00:00:00ZMedical imaging techniques for transformation and inference using deep learning(MS)
Srinivasan, Preethi
Medical imaging has significantly progressed to yeild high quality visual representations
of the organs inside the body and is of critical value to health care. Multiple imaging
modalities such as MRI, X-ray, CT and Ultrasound exist to serve different diagnostic purposes.
Nevertheless, in addition to the actual energy signal, post-processing approaches
exists which are intended to assist the diagnosis by performing simple improvements like
enhancing the sharpness and reducing image noise, providing intelligent suggestions by
segmenting the artefacts, classifying the diseases, making meaningful and critical inferences
and enabling mass screening. These post-processing tasks can be improved and yield
benefits such as decreased acquisition time, cost, need for expert training, increased comfort, and decreased radiation hazard. In this thesis, we have explored deep learning-based techniques for some advanced post-processing tasks like synthesise MR images, automate
X-ray report generation, and denoise the CT Scan. Synthesising inter modality images of MRI: MRI imaging can be utilised to interpret the distinct nature of tissues, characterised by two relaxation times, namely T1 and T2, producing contrasting yet related information. In order to reduce the acquisition time and thereby alleviate comfort and reduce the per-person cost, we propose an Encoder-Decoder based deep learning architecture to reconstruct T2 weighted image from T1 weighted image.
Automating X-ray report generation: We propose an attention-based deep neural
network to generate X-ray report automatically. X-rays can be used for mass screening in
several critical/pandemic scenarios as is fast and cost effective.
Denoising low dose CT Scan: Computed Tomography (CT) scanners induce X-ray
radiation through the body to capture images of the bones and tissues. A higher radiation
dosage leads to clearer images but have harmful effects. We propose an architecture that
computes visual attention across non-overlapping patches to denoise the low dose CT scans.
2021-02-01T00:00:00Z