SpiNet: A Deep Neural Network for Schatten p-norm Regularized
Medical Image Reconstruction
Department of Computational and Data Sciences
Model based deep learning architectures for solving inverse problems consists
of two parts, a deep learning based denoiser and an iterative data consistency solver.
The former has either L2 norm or L1 norm enforced on it, which are convex and can
be easily minimized. This work proposes a method to enforce any p norm on the
noise prior where 0 < p ≤ 2. This is achieved by using Majorization Minimisation
algorithm, which upper bounds the cost function with a convex function, thus can
be easily minimised. The proposed SpiNet has the capability to work for a fixed p
or it can learn p based on the data. The network was tested for solving the inverse
problem of reconstructing magnetic resonance (MR) images from undersampled k
space data and the results were compared with a popular model based deep learning
architecture MoDL which enforces L2 norm along with other compressive sensing
based algorithms. This comparison between MoDL and proposed SpiNet was performed for undersampling rates (R) of 2×, 4×, 6×, 8×, 12×, 16× and 20×. Multiple
figures of merit such as PSNR, SSIM and NRMSE were utilized in this comparison.
A two tailed t-test was performed for all undersampling rates and for all metrices
for proving the superior performance of proposed SpiNet compared to MoDL. For
training and testing, the same dataset that was utilized in MoDL implementation
was deployed. We also validated our algorithm on chest and dyanmic breast MRI
data.
Learning non linear mappings from data with applications to
priority based clustering, adaptive prediction, and detection
Department of Electronic Systems Engineering
With the drive for handling big data applications, algorithms to extract and
understand the underlying relations within the data have gained momentum. In the
unsupervised learning framework, we formulate a systematic theory for clustering
data with different priority over the data attributes. We also formulate a potential
function based on a spatio-temporal metric and create hierarchical vector quantization feature maps by embedding memory structures across the feature maps to
learn the spatio temporal correlations in the data across clusters. In the supervised
learning framework, we propose an architecture for prediction and classification ofvideo sequences based on an adaptive pooling layer along with detection of out of distribution classes. We also propose a learning algorithm towards the encoding of
linear block codes over binary fields.
Adaptive Sample Selection for Robust Learning under Label Noise
Department of Electrical Engineering
Deep Neural Networks have been shown to be susceptible to overfitting in the
presence of noisily-labelled data. For the problem of robust learning under such
noisy data, several algorithms have been proposed. A prominent class of algorithms
relies on sample selection strategies. For example, many algorithms use the ‘small
loss’ trick wherein a fraction of samples with loss values below a certain threshold
are selected for training. These algorithms are sensitive to such thresholds which
often depend on (typically unknown) label noise rates, and it’s difficult to fix or
learn these thresholds. We propose a data-dependent, adaptive sample selection
strategy that relies only on batch statistics of given mini-batch for robustness against
label noise. The algorithm doesn’t have any additional hyperparameters, doesn’t
need information on noise rates, and doesn’t need access to extra data with clean
labels. We empirically demonstrate the effectiveness of our algorithm on benchmark
datasets.
Independent Sets in Semi-random Hypergraphs
Department of Computer Science and Automation
In this work, we study the independent set problem on hypergraphs in a natural
semi-random family of instances. Our semi-random model is inspired by the
Feige Kilian model [FK01] This popular model has also been studied in the works of [FK01,
Ste17, MMT20] etc. McKenzie, Mehta, and Trevisan [MMT20] gave algorithms for
computing independent sets in such a semi-random family of graphs. The algorithms
by McKenzie et al. [MMT20] are based on rounding a “crude-SDP”. We generalize
their results and techniques to hypergraphs for an analogous family of hypergraph
instances. Our algorithms are based on rounding the “crude-SDP” of McKenzie et al.
[MMT20], augmented with “Lasserre/SoS like” hierarchy of constraints. Analogous
to the results of McKenzie et al. [MMT20], we study the ranges of input parameters
where we can recover the planted independent set or a large independent set.
A provably optimal algorithm for finding fair communities
Department of Computer Science and Automation
TBD
ASR in Low-resource Languages Using Domain Adaptation
Approaches
Department of Electrical Engineering
Building an automatic speech recognition (ASR) system from scratch requires
a large amount of annotated data. For many of the languages in the world, there
are not enough annotated data to train an ASR. However, there are cases where
the low-resource language shares a common acoustic space with a high resource
language, which has enough annotated data to build an ASR. In such cases, weshow that the domain independent acoustic models learnt from the high-resource
language through deep domain adaptation approaches can perform well in the ASR
task on the low-resource language. The assumption is that the low-resource language has limited audio data for domain training and good enough text data to
build a reasonable language model. We use the specific example of Hindi in the
source domain and Sanskrit in the target domain. The results suggest that domain
adaptation approaches can help in the accelerated development of ASR systems for
low-resource languages.
Speaker Diarization for conversational speech using graph clustering
Department of Electrical Engineering
Speaker diarization is the task of automatic segmentation of the given audio
recording into regions corresponding to different speakers. It is an important step
in information extraction from conversational speech. The applications range from
rich speech transcription to analysing turn-taking behavior in clinical diagnosis.
The state-of-the-art approach involves extraction of short segment features called
as speaker embeddings followed by computation of pairwise similarity scores which
are then used to perform clustering using similarity scores. In this talk,
In this talk, I will discuss various challenges and popular clustering approaches in this field, followed by
our recent work on representation learning and graph based path integral clustering
(PIC). The representation learning is based on principles of self-supervised learning.
It uses the cluster targets from PIC and the clustering step is performed on
embeddings learned from the self-supervised deep model. Our results on telephone
and meeting dataset shows improvements over the baseline approach.
Neural PLDA Modeling for End-to-End Automatic Speaker Verification
Department of Electrical Engineering
Given an enrollment recording of a particular target speaker, automatic speaker
verification (ASV) is the problem of determining whether a test speech segment is
spoken by the enrolled target speaker or not. The state-of-the-art methods used
for ASV involve an embedding extractor that takes a speech segment of arbitrary
duration and transforms it into an embedding vector of fixed dimensions. As the
embedding extractors are typically trained with classification objective, a separate
back-end model is trained to determine if a pair of speech segment embeddings
belong to the same speaker or a different speaker. In this talk, I will briefly discuss
the popular approaches for Embedding extraction and back- end modeling, followed
by our recent work on Neural Probabilistic Linear Discriminant Analysis (NPLDA)
back-end using a verification cost function. I will then discuss a simple extension to
train the embedding extractor and the NPLDA in an end-to-end framework
Simultaneous Identification of Gulf Stream and Rings from Concurrent
Satellite Images of Sea Surface Temperature and Height Images
Department of Computational and Data Sciences
Accurate digitization of synoptic ocean features is crucial for climate studies
and the operational forecasting of ocean and coupled ocean-atmosphere systems. For
operational regional models of the north Atlantic, skilled human operators visualize
and extract the Gulf Stream and Rings (warm and cold eddies) through a timeconsuming manual process. To automate this task, we develop a dynamics-inspired
deep learning system(W-Net) that extracts the Gulf Stream and Rings from concurrent satellite images of sea surface temperature (SST) and sea surface height (SSH).
Our approach’s novelty is that the above extraction task is posed as a multi label
semantic image segmentation problem solved by developing and applying a deep convolutional neural network with two parallel Encoder-Decoder networks (one branchfor SST and the other for SSH). We obtain 82.7% raw test accuracy for Gulf Stream
and more than 70% raw eddy detection accuracy. Detailed ablation studies and an
examination of both SST and SSH parts of the network is also performed.
Functional Connectivity based EEG Biometrics
Department of Electrical Engineering
Resting state brain electrical activity has evidence of unique neural signatures
motivating investigation of the discriminative information for person identification
systems. In this talk, methods are discussed to extract features from high density
electroencephalogram (EEG) for potential use in biometric applications. Functional
connectivity (FC) metrics and graph based (GB) metrics derived from FC are used
as features in a support vector machine classifier. 184 subjects pooled from 3
different datasets obtained using different protocols and acquisition systems are used
to investigate the performance of FC and GB metrics. Independent datasets are
assessed for obtaining a consistent feature set and frequency band. An identification
accuracy of 97.4% is obtained using phase locking value based measures extracted
from the gamma oscillations of the EEG data. Increase in classification accuracy
with higher frequency bands indicates the presence of unique neural signatures in
those bands.
Reliable sleep stage classification method using multiple EEG features
and RUSBoost technique
Department of Electrical Engineering
Sleep is extremely crucial for our physical and mental health, and deprivation or
poor quality of sleep can lead to numerous health problems. Sleep patterns can also
be used to diagnose neurological disorders like dementia, Alzheimer’s etc. Therefore,
it is important to analyze the sleep signals accurately. Traditionally, sleep scoring
is done by the sleep experts by visual inspection of overnight polysomnography
(PSG) data. However, this manual scoring suffers from issues like subjectivity,
high cost, time-consuming and inter-rater variability. An automatic sleep scoring
system can be useful to tackle these problems. This work proposes a reliable sleep
scoring method by using a fusion of multiple EEG features and an ensemble classifier
called random undersampling with boosting technique (RUSBoost). The results
achieved with a single channel EEG are comparable or better than the
state-of-the-art methods in the literature for both two-class (sleep-awake) and multi-class
classifications, on three different publicly available sleep databases. The proposed
method also provided promising results on the unseen test subjects using both 50%-
holdout as well as 10-fold cross-validation approach, emphasizing the reliability of
the method.
Automatic speech recognition for speech conversations
Department of Electrical Engineering
The transcription of long-form conversations between multiple speakers is
challenging, due to the speaker variability, unknown segment end-points and the
conversational language. Further, the speakers may speak simultaneously, leading to
overlapped speech in single channel recordings. A “rich transcription” is often desired
for downstream applications in speech analytics, recording archival, audio
indexing, etc. In this talk, we discuss a novel approach for the tran- scription of speech
conversations with natural speaker overlap, from single channel recordings.
We propose a combination of a speaker diarization system and a hybrid automatic speech
recognition (ASR) system with speaker activity assisted acoustic model (AM). An
end-to-end neural network system is used for speaker diarization, and the acoustic
model is trained to output speaker specific senones. Further, we propose
enhancements to language model using conversation context. We illustrate the benefits
of the proposed multi-speaker ASR system for speech conversations using natural
telephone conversations from Switchboard corpus.
CDW induced phase transition in layered heterostructures
Department of Electrical Communication Engineering
1T-TaS2 is a unique layered material exhibiting strong charge density wave
(CDW) driven resistance switching that can be controlled by temperature, an external electric field as well as optical pulses. However, such resistivity switching
effects are often weak, and cannot be modulated by an external gate voltage – limiting their widespread usage. Using a back-gated 1T-TaS2/2H-MoS2 heterojunction,we achieve a gate-controlled resistivity switching during the nearly-commensurate
(NC) to incommensurate (IC) phase transition of TaS2, and the switching ratio is
enhanced by a factor of approximately 14.5 compared to the standalone TaS2 control. This will
boost a plethora of device applications that exploit these phase transitions, such
as ultra-broadband photodetection, negative differential conductance, fast oscillator
and threshold switching in neuromorphic circuits. Interestingly, by simply changing
the geomertry we demonstrate a negative differential resistor (NDR) in an asymmetrically designed 1T-TaS2/2H-MoS2 T-junction using an electrically driven CDW
phase transition.
Excitons in flatlands
Department of Electrical Communication Engineering
An exciton is a bound state of an electron and an electron hole which are
attracted to each other by the electrostatic Coulomb force. When the dimensionality
of the crystals reduced to atomic scale, the binding energy of the excitons enhance
dramatically due reduced dielectric screening and increased quantum confinement.
The larger binding energy in these low dimensional semiconductors persist even
in multilayers compared to conventional semiconductors. In fact, in a multilayer
system the excitons can form an even-odd type alternating pair and retain the 2d
properties. In a bilayer system the excitons can be completely confined to a single
layer (intralayer exciton) or the electron and hole can be in different layer (interlayer
exciton). With an application of vertical electric field, the intralayer excitons can
be converted to interlayer and the corresponding exciton oscillator strength can be
significantly modulated. Further, by changing the carrier concentration the oscillator
strength of excitons can also be tuned corresponding to the trions
Theory of nonvolatile resistive switching in monolayer molybdenum
disulfide with passive electrodes
Department of Electronic Systems Engineering
Resistive-memory devices promise to revolutionize modern computer
architecture eliminating the data- shuttling bottleneck between memory
and processing unit.Recent years have seen a surge of experimental demonstrations of 2D material-based
metal-insulator-metal devices. However, the fundamental mechanism of switching
has remained elusive. Here, we conduct reactive molecular dynamics simulations
for sulfur vacancy inhabited monolayer molybdenum disulfide-based device to gain
insight into such phenomena. We observe that with the application of a suitable
electric field, at the vacancy positions, the sulfur atom from the other plane pops and
gets arrested in the plane of the molybdenum atoms. Rigorous first-principles
calculations surprisingly reveal localized metallic states (virtual filament) and stronger
chemical bonding for this new atomic arrangement, explaining nonvolatile resistive
switching. We further observe that localized Joule heating plays a crucial role in
restoring the original position of popped sulfur atom. The proposed theory may
provide useful guidelines for designing high-performance resistive-memory-based computing architecture.
Learning Based Classification of Magnetism in 2D materia
Department of Electronic System Engineering
Predicting magnetism in two-dimensional materials with fundamental
information about the crystal is a challenging task in the field of material science. So
far, the magnetism in two-dimensional materials is being predicted computationally.
With the popularity of machine learning algorithms in predicting material
properties with accuracy close to ab initio calculations and computational speed order of
magnitude faster, we have tried to use machine learning for predicting magnetism
in two-dimensional materials. With a very small database of these materials, alongwith
lots of boundary cases and sparsity in data, make it a difficult task for the
algorithms to build a classifier that can classify such imbalanced overlapped datasets.
We have tried exploring the available crystal graph network for predicting
magnetism. Here, we use basic information of crystal structure like atomic and bond
features to design a model that can predict magnetism in such materials with good
accuracy
An Evaluation of Methods to Port Legacy Code to SGX Enclaves
Department of Computer Science and Automation
The Intel Security Guard Extensions (SGX) architecture enables the abstraction of enclaved execution, using which an application can protect its code and
data from powerful adversaries, including system software that executes with the
highest processor privilege. While the Intel SGX architecture exports an ISA withlow-level instructions that enable applications to create enclaves, the task of writing
applications using this ISA has been left to the software community.
We consider the problem of porting legacy applications to SGX enclaves. In
the approximately four years to date since the Intel SGX became commercially
available, the community has developed three different models to port applications to
enclaves—the library OS, the library wrapper, and the instruction wrapper models.
In this paper, we conduct an empirical evaluation of the merits and costs of
each model. We report on our attempt to port a handful of real-world application
benchmarks (including OpenSSL, Memcached, a Web server and a Python interpreter) to SGX enclaves using prototypes that embody each of the above models.
Our evaluation focuses on the merits and costs of each of these models from the
perspective of the effort required to port code under each of these models, the effort
to re-engineer an application to work with enclaves, the security offered by each
model, and the runtime performance of the applications under these models.
Non-malleable Codes
Department of Computer Science and Automation and Department of
Mathematics
Non-malleable codes are coding schemes that help in securing cryptographic
protocols against a class of attacks called as “related-key attacks”, which allow the
adversary to tamper with the secret key (on which the security of the protocol relies)
and observe additional input-output behaviour of the protocol on the tampered key,
leading to a security breach. Non-malleable codes, informally, give a guarantee
that, under a tampering attack, the recovered tampered message is either same as
the original message, or is independent of it.
In this talk, I will be formally defining these primitives and further talk about my
work towards building such codes for a strong adversarial setting with optimal rate.
Furthermore, I would also talk about some applications of non-malleable codes to
other interesting cryptographic primitives
Heap Aware Symbolic Execution
Department of Computer Science and Automation
In this paper, we show that dynamic symbolilc execution, when coupled
with a well-defined model of the allocator behavior, can be effective in detecting
heap based vulnerabilities. Specifically, we model the behavior of heap memory
allocator to avoid states that will not occur in a concrete execution and define
precise conditions that represent heap based vulnerabilities by extracting metadata
from the allocator. We develop a tool named RADAR that uses these techniques in
tandem on ptmalloc memory allocator. RADAR effectively detects vulnerabilities in
a set of micro-benchmarks and real-world projects fetched from github.
Scaling Blockchains Using Coding Theory and Verifiable Computing
Department of Computer Science and Automations
The issue of scalability has been restricting blockchains from their widespread
adoption. In this talk, we will focus on blockchain scaling solutions inspired by
coding theory. We first consider SeF, a blockchain archiving architecture using LT
codes to reduce storage constraints by 1000x. SeF enables full nodes to store a small
number of encoded blocks or droplets instead of a complete blockchain. While other
rate-less codes utilizing two encoding levels are proven better than LT codes, we
investigate their suitability in the proposed architecture. We propose and simulate
three techniques about how to incorporate these coding strategies.
The other work we examine is PolyShard, which introduces the notion of codedsharding, making sharding resilient to an adaptive adversary. However innovative,
PolyShard requires decoding RS codes over large fields, making it computationally
intensive and less practical. We propose replacing the decoding phase with verifiable
computing, reducing the bottleneck, and making architecture practical for light
verification functions.
MPC for small population with applications to Privacy-Preserving
Machine Learning
Department of Computer Science and Automation
Performing machine learning (ML) computation on private data while maintaining data privacy, aka Privacy preserving Machine Learning (PPML), is an emergent field of research. Recently, PPML has seen a visible shift towards adopting the
Secure Outsourced Computation (SOC) paradigm due to the heavy computation.
In the SOC paradigm, computation is outsourced to a set of powerful and specially
equipped servers that provide service on a pay per use basis. In this work, we propose a robust PPML framework using secure multi party computation (MPC) for
a range of ML algorithms in the SOC setting that guarantees output delivery to
the users irrespective of any adversarial behaviour. Robustness, a highly desirable
feature, evokes user participation without the fear of denial of service.
We demonstrate our framework’s practical relevance by benchmarking popular ML algorithms such as Logistic Regression and deep Neural Networks such as
VGG16 and LeNet, both over a 64-bit ring in a WAN setting.
Towards Deep Robust Deconvolution of Low-Dose Perfusion Computed
Tomography
Department of Computational and Data Sciences
Computed Tomography Perfusion (CTP) has established itself as an economic,
fast and reliable clinical modality for diagnosis of cerebrovascular diseases such as
acute ischemia and vasospasm. To obtain the necessary perfusion maps for accurate
diagnosis, it is mandatory to subject the patient to high radiation dosage, which
in turn have several pernicious biological side effects. Thus, obtaining high-quality
perfusion maps even at low radiation dosage remains an interesting and convoluted
challenge. To combat the same, we propose a novel deep learning based end-to-end
framework, to produce high quality Cerebral Blood Flow (CBF) maps from low-dose
raw CTP data. The proposed model is able to perform the deconvolution without
explicit information on the Arterial Input Function (AIF) and isn’t susceptible to
varying levels of noise. Our experiments and results validate the superiority of our
framework over existing state of the art algorithms.
Wavelet Design in a Learning Framework
National Mathematics Initiative & Department of Electrical Engineering
The wavelet transform has proven to be a highly successful data representation tool, and is used in several signal and image processing applications,
such as compression, denoising, singularity detection, etc. The problem of wavelet
design has traditionally been approached from an analytical perspective, using tools
from harmonic analysis to construct continuous-domain functions satisfying the admissibility criterion. In this talk, a learning based approach to wavelet design is
introduced, which provides a coherent computational framework for addressing the
design problem. Perfect reconstruction filterbanks are viewed as convolutional autoencoders, and wavelets are designed by training these filterbank autoencoders. The
designed wavelets are data-independent, which precludes the need for customized
datasets. In fact, high-dimensional Gaussian data vectors can be used for training,
and it will be shown that a near-zero training loss implies that the learnt filters
satisfy the perfect reconstruction property with very high probability. Properties
of a wavelet such as orthogonality, compact support, smoothness, symmetry, and
vanishing moments are incorporated by designing the autoencoder architecture appropriately, with a suitable regularization term added to the mean-squared error cost
when needed. The proposed approach not only recovers the well known Daubechies
family of orthogonal wavelets and the Cohen-Daubechies-Feauveau family of symmetric biorthogonal wavelets, but also learns wavelets outside these families.
Appearance Consensus Driven Self-Supervised Human Mesh Recovery
Department of Computational and Data Sciences
We present a self-supervised human mesh recovery framework to infer human
pose and shape from monocular images in the absence of any paired supervision.
Almost all the avalble approaches aims to regress a parametric human model by
supervising on datasets with 2D landmark annotations. Different from these, we
propose a novel appearance consensus driven self-supervised objective. To effectively
disentangle the foreground (FG) human we rely on image pairs depicting the same
person (consistent FG) in varied pose and background (BG) which are obtained from
unlabeled wild videos. The proposed FG appearance consistency objective makes
use of a novel, differentiable color-recovery module to obtain vertex colors without
the need for any appearance network, via efficient realization of color-picking and
reflectional symmetry. Furthermore, the resulting colored mesh prediction opens up
the usage of our framework for a variety of appearance-related tasks beyond pose
and shape estimation.
Nonconvex Analysis-Sparse Model — Moreau Envelope Regularization
and Projected Proximal Gradient Descent Optimization
Department of Electrical Engineering
We consider the problem of nonconvex analysis-sparse recovery in which the
signal is assumed to be sparse in a highly redundant analysis operator. Standard
non-convex sparsity promoting priors do not have a proximal operator in closed
form under a redundant analysis operator and therefore, proximal approaches
cannot be applied directly. This motivated us to develop two alternatives – Moreau
envelope regularization and projected transformation for solving the problem under
con- sideration. First, we consider the Moreau envelope counterpart to nonconvex
penalty defined in the analysis setting as a sparsity enforcing prior and develop
algorithms for optimizing the corre- sponding cost. Second, we employ a projection
transform that maps the analysis-sparse recovery problem into an equivalent
constrained synthesis-sparse formulation, which can be optimized by the projected
proximal algorithm. Finally, we demonstrate the efficacy of the proposed techniques
in comparison with benchmark techniques in a real application, namely, 2-D
magnetic resonance imaging (MRI) reconstruction considering shift-invariant discrete
wavelet transform (SIDWT) as the redundant analysis operator.
Zero-shot Sketch-based Image Retrieva
Department of Electrical Engineering
Sketch-based image retrieval (SBIR) addresses the problem of ranking images
in a database according to their relevance to a given hand-drawn sketch-query. It is
an important research area, though earlier SBIR protocol assumed that the query
always belongs to one of the training categories. Recently an evolved version,
Zero shot SBIR (ZS-SBIR), has emerged; which breaks down this restriction and thus
the query may belong to categories, either seen or unseen to the model. The goal of
my doctoral research is addressing different challenges, i.e. the assumption of paired
sketch-image training data, attribute-augmented query etc. associated with
ZSSBIR. The focus of this talk is to discuss one such challenge, the effect of imbalance
in training data. Towards this goal, we proposed an adaptive-margin based
regularizer which can be seamlessly incorporated with any existing ZS-SBIR algorithm to
improve its performance under imbalanced data. We discuss relevant experiments
to demonstrate its effectiveness.
Siamese-SR: A Siamese Super-Resolution model for boosting resolution
of Digital Rock images for improved petrophysical property estimation
Department of Computational and Data Sciences
The Digital Rock workflow is an emerging framework utilizing advances in
imaging technologies and state-of-the-art image processing algorithms to construct
digital models of reservoir rocks, which are used to run physics simulations to
calculate several petrophysical properties of interest. The accuracy of the digital rock
workflow relies crucially on the resolution of the acquired images which is currently
limited by the hardware of the micro-CT scanning technology. Super-resolutionmethods
can effectively address this limitation by digitally boosting the resolution
of images acquired using micro-CT scanners. We have proposed a novel deep
learning based super-resolution model called Siamese-SR to improve the resolution of
Digital Rock images whilst retaining the texture and providing optimal denoising.
The Siamese-SR model consists of a generator which is adversarially trained with
a Relativistic and a Siamese Discriminator utilizing Materials In Context (MINC)
loss estimator. We also propose to move away from image-based metrics (SSIM &
PSNR) to physics based quantification metrics which are based on running
subsequent steps in the workflow mainly calculation of image based porosity and running
Mercury Injection Capillary Pressure (MICP) simulations for calculation of accurate
petrophysical parameters of interest.
Revealing Disocclusions in Temporal View Synthesis for Frame Rate
Upsampling
Department of Electrical Communication Engineering
Suppose a user is exploring a virtual environment on a head mounted display
device. Given the past view of a rendered video frame and the most recent head
position, our goal is to directly predict a future video frame without having to
graphically render it. We refer to the above problem as egomotion-aware temporal
view synthesis. The biggest challenge in such view synthesis through warping is the
disocclusion of the background. For infilling them, we introduce the idea of infilling
vectors which point from disoccluded regions to known regions in the synthesized
view. Tapping the correlation between camera motion in the past frames, we design
a learning framework with temporal guidance to predict these infilling vectors. Our
extensive experiments on a large scale dataset that we build for evaluating temporal
view synthesis, in addition to the SceneNet RGB-D dataset, demonstrate that our
infilling vector prediction approach achieves superior quantitative and qualitative
infilling performance compared to other approaches in literature.
Sparsity Driven Latent Space Sampling for Generative Prior based
Compressive Sensing
Department of Electrical Engineering
We address the problem of recovering signals from compressed
measurements based on generative priors. Recently, generative model based compressive
sensing (GMCS) methods have shown superior performance over traditional
compressive sensing (CS) techniques in recovering signals from fewer measurements.
However, it is possible to further improve the performance of GMCS by introducing
controlled sparsity in the latent-space. We propose a proximal meta-learning (PML)
algorithm to enforce sparsity in the latent space while training. Sparsity naturally
divides the latent space and leads to a union-of-submanifolds model in the solution
space. The overall framework is named as sparsity driven latent space sampling
(SDLSS). In addition, we derive the sample complexity bounds for the proposed
model. Furthermore, we demonstrate the efficacy of the proposed framework over
the state-of-the-art techniques with application to CS on standard datasets such as
MNIST and CIFAR-10. Our findings show that the proposed approach improves
the accuracy and aids in faster recovery compared to GMCS.
There is More to a Spectrogram Than Meets the Eye!
Department of Electrical Engineering
Developing methods for accurate speech analysis has a direct impact on applications such as speech synthesis, speaker recognition, speech recognition, and
voice morphing, etc. A widely used tool to analyze speech is the spectrogram,
which represents the time-varying spectral content of the speech signal. We model
a spectrotemporal patch of the spectrogram by using 2-D amplitude-modulated and
frequency-modulated (AM-FM) sinusoids and solve the demodulation problem in
2-D using a novel tool, namely, the Riesz transform. The AM and FM componentscorrespond to the vocal tract smooth envelope and excitation signal. We demonstrate the impact of the Riesz transform technique for applications such as vocal
tract filter estimation, voiced/unvoiced separation, pitch tracking, and aperiodicity
estimation. The effectiveness of the new representation and speech parameters is
shown for the task of speech reconstruction using WaveNet, a neural vocoder.
Vital signs monitoring using FMCW radar
Department of Electrical Engineering
In this talk, the discussion is on non-contact vital signs monitoring using a
frequency modulated continuous wave (FMCW) radar. Here, we analyse phase of
the Intermediate Frequency signal (IF) in detail. Any human or clutters which is
exposed to FMCW radar signal is able to reflect the chirp, the reflected chirp goes
to mixer to produce IF signal. This paper deals with extracting vital signs such asbreathing rate, heart rate from a subject who sits in front of radar irrespective of the
chest orientation towards the radar. Phase of the IF signal is processed to obtain
the spectrum in desired range-bin. The spectrum is further subjected to filtering
and threshold as will be detailed. Random body movements are detected and those
samples are discarded for calculation of heart-rate. We also classify the subjects
based on breathe apnea using avaialble data set.
Resource Allocation with Limited Feedback in Underlay D2D Networks
Department of Electrical Communication Engineering
Device to device (D2D) communication enables direct communication between
the devices. In underlay D2D, D2D users reuse the subchannels assigned to cellular
users (CUs), which leads to interference between them. We explore the problem of
allocating multiple D2D pairs per subchannel and the resulting trade off between the
spatial reuse and inter-D2D interference. We study the practical partial and statistical channel state information (CSI) models in which the D2D user only knows the statistics of inter D2D and inter cell interferences. We propose a limited feedback
scheme, where the D2D user computes its rate with a reliability guarantee despite
the limited CSI and feeds it back to the base station. We propose interference-aware
resource allocation algorithms that allocate multiple D2D pairs per subchannel to
maximize D2D sum rate while guaranteeing minimum rate for the CUs. The proposed algorithms provably achieve at least one third and half of the optimal sum
rate. Numerically, they achieve near-optimal performance.
Design of k-space trajectories to reduce scan time for an MRI system
Department of Electrical Communication Engineering
MRI is an essential non-invasive medical imaging modality. However, it is
limited by long scan times. The sampling in MRI happens in the 2D/3D Fourier
domain called as the k-space. It is traversed along continuous paths called
trajectories. One of the more recent ways to reduce scan time in MRI is using compressed
sensing (CS) methods which allow random undersampling of the k-space. However,
this does not satisfy system constraints of MRI. Here, the design of feasible
trajectories to traverse the k-space is discussed. We propose a generalized framework
that encompasses projection-based methods to generate feasible trajectories. This
framework allows to construct feasible trajectories from both random and
structured initial trajectories, e.g., based on the traveling salesman problem (TSP). It is
observed that the proposed TSP-based and random-like trajectories provide better
reconstruction performance than the state-of-the art methods with a shorter scan
time. In particular, the SIP-random method improves reconstruction performance
with a 26% reduction in scan time
Probabilistic Forwarding with Coding for Network-wide Broadcast
Department of Electrical Communication Engineering
Broadcasting information is crucial in ad-hoc networks such as Internet
of Things and Wireless Sensor Networks which have no centralized infrastructure.
Mechanisms such as flooding are not suitable as they involve redundant
transmissions. In our work, we propose and analyze a completely distributed, low complexity,
energy-efficient broadcast algorithm for these networks.
We consider a network with a distinguished source node that has data packets
to broadcast. It encodes these data packets into n coded packets in such a way that,
any node receiving at least k out of the n coded packets can retrieve the original data
packets. The source transmits the n coded packets to its one-hop neighbours. Every
other node in the network follows a probabilistic forwarding protocol: it forwards
a previously unreceived packet to all its neighbours with probability p and does
nothing with probability 1 − p.
Our primary interest is to analyze this mechanism on random geometric graphs
(RGGs) which are used to model deployments of ad-hoc networks. We find that
with a judicious choice of n and the forwarding probability p, we can significantly
reduce the expected total number of transmissions needed to broadcast data from
the source. Moreover, while this holds for other well-connected graphs (grids, lattices
etc.), it is not true for trees. We provide theoretical justifications for the same
Second order Rate Bounds for a Block Fading Channel Powered
by an Energy Harvesting Transmitter
Department of ECE
In this talk, we illustrate how second-order lower and upper bounds can be
derived for a point to point complex Gaussian fading channel with an energy harvesting transmitter, and fading gains available at the transmitter and the receiver.
A standard approach to derive such results for a channel with finite states makes
use of the method of types, especially in the derivation of upper bound. Such an
approach is not feasible in our case as the number of states is uncountable. Further,
a dynamic power constraint on the codeword symbols, imposed by the energy harvesting transmitter, has to be met. We illustrate how strong approximation principle
turns in handy in deriving bounds in such a scenario, especially while deriving upper
bounds. As a special case, we obtain corresponding results for no state information
at the transmitter and full state information at the receiver.
MCS selection algorithms for URLLC in Multi-connectivity
Department of Electrical Communication Engineering
Ultra-reliable and low-latency communications, is a new field in 5G, requires
very high reliability and very low latency. It will deliver services like factory automation, telesurgery and autonomous driving. Multi-connectivity, in which multiple
base stations (BSs) send the same information to the receiver, is crucial for improving reliability. In addition, to satisfy the latency requirement, URLLC data must be
scheduled immediately, preempting conventional traffic. We propose an algorithm
that selects one or more BSs and their modulation and coding schemes (MCSs)
to minimize the degradation of the broadband data while meeting the reliability
requirements of URLLC.We then generalize our approach to account for the time-varying nature of
the wireless channel, which causes the periodic channel quality feedback reports to
become outdated. For this scenario, we propose a novel outage probability-based
definition of reliability and a new algorithm to determine the BSs and their MCSs.
Streaming Codes for Reliable, Low-Latency Communication
Department of Electrical Communication Engineering
Reliable, low-latency communication is key to many envisaged next-generation
communication applications such as assisted driving, augmented reality and telesurgery.
Streaming codes represent an approach to forward error-correction at the packet level
in which a dropped or lost packet is viewed as an erased code symbol and where the
codes are designed to be operated under a decoding-delay constraint.This talk will
focus on our recent constructions of streaming code. These
constructions are based on careful embedding of code symbols belonging to a scalar
block code within the packet stream. Different embedding approaches and attendant
constructions will be presented and compared.
Design of Reduced Feedback Scheme, User Scheduling, and MAC
protocols for Full-Duplex Communication Systems
Department of Electrical Communication Engineering
Full-duplex (FD) promises to double the data rates by enabling simultaneous
transmission and reception over the same frequency band. Enabling FD in cellular
networks creates inter- user interference between uplink and downlink. We propose
a novel user-pair scheduling and mode selection algorithm (UPSMA) that maximizes
sum spectral efficiency, and we characterize its asymptotic behaviour. We also
propose a reduced feedback scheme for it, in which each user feeds back a limited
number of inter-user interferences. We then study FD wireless local area networks
(WLANs). We propose a medium access control (MAC) protocol called asymmetric
FD-MAC (AFD-MAC). AFD-MAC intelligently uses hidden users to exploit FD.
It uses the random back-off counter-based channel contention mechanism of 802.11
and introduces two new signals to provide as many FD transmission and reception
opportunities to the AP as possible.
Scheduling and Power Control for Wireless Multicast Systems via Deep
Reinforcement Learning
Department of Electrical Communication Engineering
Multicasting in wireless systems exploits the redundancy in user requests in a
Content Centric Networks. Power control and optimal scheduling can significantly
improve the wireless multicast network’s performance under fading. Model-based
approaches for power control and scheduling are not scalable to large state space or
changing system dynamics. Here, we propose a constrained Deep RL approach to
obtain a power control policy that matches the optimal policy for a small network.
We show that power control policy can be learnt for reasonably large systems via
this approach. Further we use multi-timescale stochastic optimization to maintain
the average power. Finally, we extend the multi-time scale approach to
simultaneously learn the optimal queueing strategy along with power control. We demonstrate
scalability, tracking and cross-layer optimization capabilities of our algorithms via
simulations. The proposed approach can be used in general large state-space
dynamical systems with multiple objectives and may be of independent in
Medium refractive index contrast sub-wavelength dielectric structures
for nonlinear optics and sensing application
Department of Electrical Communication Engineering
Dielectric sub-wavelength structures are of increased interest for enhancing
light matter interaction, owing to its capability to mould the field within or aroundthe structure.
CMOS compatible materials like silicon nitride (SiN), silicon dioxide
(SiO2), having a wide transparency window ranging from visible to mid-infrared
wavelengths make suitable grating materials. They, however, lack sufficient
scattering efficiency due to its weak refractive index contrast with respect to air. In this
work, we study one-dimensional SiN based guided mode resonant (GMR) structures
for enhancing the fluorescence from Rhodamine B dye using polarization sensitive
dual mode resonances making it conducive for sensitive imaging of bio-samples with
low analyte concentration and for improving the nonlinear signals from amorphous
silicon overlayers at lower power inputs
Design and Development of a Dual Functional Intubation Catheter for
Chronic Airway Management
Department of Electronic Systems Engineering
In this work, we report the design and development of an intubation catheter
for grading stenosis in the pediatric upper airway. Severely constricted airways often
warrant continuous monitoring as resistance to flow increases to fourth power for
every one-degree reduction in tracheal patency. The complexities and impediments
with conventional diagnostic tools such as misclassification on the degree of narrowing and long radiation exposure make them sub- optimal for emergent diagnosis. We
propose an intraoperative diagnostic tool integrated with an array of flow sensorsand tactile sensors mounted at the distal end of the manoeuvrable catheter. The
flow sensors generate localised flow patterns across various segments of the tracheobronchial tree with unique peaks corresponding to the site of obstruction. While
the flow patterns allow for locating stenosis, the tactile sensors can determine the
target tissue stiffness. Quantitative evaluation of alteration in the airway column
biomechanics facilitates targeted diagnosis and expedites on-site decision making.
s
A portable MEMS-based platform for the electrothermal phenotyping of breast biopsy tissue: towards a system for tumour
delineation
Centre for BioSystems Science and Engineering
We report a portable diagnostic platform integrated with MEMS-based biochips
for rapid phenotyping of breast biopsy tissues using electro-thermal modalities. The
biochips fabricated on a silicon substrate using microfabrication techniques measure the bulk resistivity (ρB), surface resistivity (ρS), and thermal conductivity (k)
of the samples. Measurements were performed from deparaffinized and formalinfixed tumour and adjacent normal breast biopsy samples from N = 8 patients. For
formalin fixed samples, the mean ρB for tumour reported a statistically significant
fold change of 4.42 ( p = 0. 014) when the tissue was heated from 25 degree to 37 degree,
while for normal the fold change was 3.47. The mean ρS measurements also reported
similar trend. The normal tissue reported a mean k of 0.563 ± 0.028 W.m−1.K−1,
while tumour reported significantly lower k of 0.309 ± 0.02 W.m−1.K−1. These observations, after validation with fresh tissues, could potentially be used for rapid
tumour delineation.
Design and Development of a Micro-Force Sensor for Ablation Catheters
Department of Electronic Systems Engineering
In this work, we present simulation, fabrication, and characterization of a
micro force sensor that has the potential to integrate with the ablation catheters for
real-time measurement of catheter tip contact forces. FEM analysis was used to
optimize the sensor design. The sensor was charac- terized using a custom-built
micro-indentation setup and automated using LabVIEW. The sensor characteristics
such as linearity, sensitivity, and hysteresis were evaluated. The sensor showed a
linear response with R2–value of 0.995 for the force ranging from 0 –1.2 N. The
sensitivity and hys- teresis were found to be 108 ±11WN−1 and 5% respectively. The
fabricated sensor was integrated into a customized catheter tip and tested on the
porcine heart tissues for evaluating the tip contact forces
Design and development of Opto-Thermo-Acoustic (OTA) based
measurement techniques to aid breast cancer diagnosis
Department of Electronic Systems Engineering
Breast cancer is among the leading cause of mortality among women
world wide. Histopathology that accounts for the microscopic evaluation of thin slices of
breast biopsy tissue under a microscope to study cell structure and morphology is
considered the gold standard. We design and develop a system that combines the
optical, thermal, and acoustic modalities to estimate the bulk tissue optical properties
(reduced scattering and absorption coefficient), thermal properties (thermal co
nductivity and specific heat), and acoustic (attenuation coefficient). The additional
information gauged through these measurements can potentially aid the pathologist
in confirming the breast cancer diagnosis. The measurements were performed on
formalin-fixed (FF) and deparaffinized (DP) breast biopsy tissues from a total of N
= 24 subjects.
Towards unique device identification using Physically Unclonable
Functions
Department of Electronic Systems Engineering
Uniqueness of device-ID is critical to ensure trustworthiness of the data
reported by the device. This work presents a novel device-specific identifier - IoT-ID,
based on Physically Unclonable Functions (PUFs) that exploits variations in
thesemiconductor manufacturing process. We design novel PUFs for Commercially Off
the Shelf (COTS) components such as clock oscillators and ADC, to derive IoT-ID
for a device. Hitherto, system component PUFs are either invasive or rely on
additional dedicated hardware circuitry to create a unique fingerprint. A highlight of
our PUFs is doing away with special hardware. IoT-ID is non-invasive and can be
invoked using simple software APIs running on COTS components. During our
evaluation, IoT-ID has demonstrated 100% accuracy in uniquely identifying 50 devices
in our deployment. We show the scalability of IoT-ID with the help of numerical
analysis on 1000s of IoT devices.
Estimation of stochastic responses due to multiple variabilities in
electromagnetic finite element modelling
Department of Computational and Data Sciences
Unmanned Aerial Vehicles (UAVs), or drones, are increasingly used for urban
applications like traffic monitoring and construction surveys. A common activity is
to hover and observe a location using on-board cameras. Advances in Deep Neural
Networks (DNNs) allow such videos to be analyzed for automated decision making.
UAVs also host edge computing capability for on-board inferencing by such DNNs.
Here, we propose a novel Mission Scheduling Problem (MSP) for co-scheduling the
flight route to visit and record video at waypoints, and their subsequent on-board
analysis, for a fleet of drones. The schedule maximizes the utility from the activities,
while meeting activity deadlines, and the energy and computing constraints. We
provide two efficient heuristic algorithms, JSC and VRC, to obtain fast, sub-optimal
solutions. Our detailed evaluation of these algorithms using real drone benchmarks
show the utility-runtime trade-offs of the 3 schedulers under diverse workloads.
Optimal Link Capacity Selection and Edge Server Placement using
Column Generation
Department of Electronic Systems Engineering
When an end user wants to subscribe to a gaming service, he expects the
network to respond quickly and provide good Quality of Service and Quality of
Experience. Any network has few Data Centers (DC), and they are placed deep
inside the core network. Therefore, it becomes challenging to provide a service to
a user which requires high bandwidth and very low latency. Placing servers near
the user and on the edge of the core network can help overcome this challenge.
Thus, the question of optimal placement of these servers arises. Also, to reduce
the total cost incurred, it is essential to activate only those links which connect
users to the DC and assign appropriate capacities to these links. We present the
Network Design Problem of optimal server placement and link capacity selection.
The problem formulated is a Large Integer Linear Program and we provide exact
and fast solution by using the Column Generation (CG) Technique.
Automatically finding bugs in web applications
Department of CSA
Most modern web applications are based on the Model view controller design
pattern. It is desirable to detect bugs automatically in controllers based on specifications given by developers. We propose an approach for this problem. The core
step of our approach is to automatically extract a formal model of a controller using
program analysis techniques. The model is based on relational algebra, and represents the database accesses and queries performed by the controller. We then feedthis model to a pre-existing constraint solver along with the user’s specification. We
have implemented our approach as a tool. We evaluated our tool on a set of 56
specifications. The tool found 34 of these to be satisfied; of the rest, upon manual
analysis, we found that two were genuinely violated, while the remaining 20 ‘unsatisfied’ warnings were actually false positives. Apart from verification purposes our
model can also aid in developer understanding.
Hardware Accelerator For Particle Transport Simulation
Department of Electronic Systems Engineering
Simulating Particle Transport through matter lies at the heart of Particle
Physics applications, including Nuclear and High Energy Physics, extending to
medical physics for treatment planning and radiation dose calculation. Our current work
investigates the principle of particle transport simulation, identifying major
computational elements involved and developing an efficient architecture for accelerating
the computation on hardware. The ultimate goal is to implement the system on a
Field Programmable Gate Array (FPGA). For a non-physicist, it is cumbersome to
develop such a system from scratch. We show how we can leverage the power of
High-Level Synthesis Tools to directly convert existing C/C++ Code for Particle
Transport to hardware implementation, highlighting the benefits and downsides of
such an approach. The final System Architecture is based on the Xilinx Vitis tool
targeted to the Xilinx Alveo U250 FPGA card. The execution model for the same
will also be discussed
Optimal Algorithms for Range Searching over Multi-Armed Bandits
Department of Computer Science and Automation
This paper studies a multi-armed bandit (MAB) version of the range-searching
problem. In its basic form, range searching considers as input a set of points (on
the real line) and a collection of (real) intervals. Here, with each specified point, we
have an associated weight, and the problem objective is to find a maximum-weight
point within every given interval.
The current work addresses range searching with stochastic weights: each point
corresponds to an arm (that admits sample access) and the point’s weight is the
(unknown) mean of the underlying distribution. In this MAB setup, we develop
sample efficient algorithms that find, with high probability, near-optimal arms within the
given intervals, i.e., we obtain PAC (probably approximately correct) guarantees.
We also provide an algorithm for a generalization wherein the weight of each point
is a multi-dimensional vector. The sample complexities of our algorithms depend,
in particular, on the size of the optimal hitting set of the given intervals.
Finally, we establish lower bounds proving that the obtained sample
complexities are essentially tight. Our results highlight the significance of geometric
constructs—specifically, hitting sets—in our MAB setting.
Tracking an AR(1) Process with limited communication per unit time
Department of Robert Bosch Centre for Cyberphysical Systems
Samples from a high-dimensional AR[1] process are quantized and sent over a
communication channel of finite capacity. The receiver seeks to form an estimate
of the process in real-time. We consider a time-slotted communication model in
slow-sampling regime where multiple communication slots occur between two
sampling instants. We propose a successive update scheme which uses communication
between sampling instants to refine estimates of the latest sample. We study the
following question: Is it better to form refined estimates and send them over
multiple communication slots, making the receiver wait more for an update, or to be fast
but loose and send new information in every communication opportunity? We show
that the fast but loose successive update scheme with spherical codes is universally
optimal asymptotically for large dimension.
Translating Natural Language Instructions to Computer Programs for
Robot Manipulation
Department of Electrical Communication Engineering
It is highly desirable for robots that work alongside humans to be able to
understand instructions in natural language. Existing language conditioned imitation
learning models directly predict the actuator commands from the image observation
and the instruction text. Rather than directly predicting actuator commands, we
propose translating the natural language instruction to a Python function which
queries the scene by accessing the output of the object detector and controls the
robot to perform the specified task. This enables the use of non-differentiable
modules such as a constraint solver when computing commands to the robot. Moreover,
the labels in this setup are significantly more informative computer programs that
capture the intent of the expert rather than teleoperated demonstrations. We show
that the proposed method performs better than training a neural network to directly
predict the robot actions
Latency Redundancy Tradeoff in Distributed Read-Write Systems
Department of Electrical Communication Engineering
Data is replicated and stored redundantly over multiple servers for availability
in distributed databases. We focus on databases with frequent reads and writes,
where both read and write latencies are important. This is in contrast to databases
designed primarily for either read or write applications. Redundancy has contrasting
effects on read and write latency. Read latency can be reduced by potential parallel
access from multiple servers, whereas write latency increases as a larger number
of replicas have to be updated. We quantify this tradeoff between read and write
latency as a function of redundancy, and provide a closed-form approximation when
the request arrival is Poisson and the service is memoryless. We empirically show
that this approximation is tight across all ranges of system parameters. Thus, we
provide guidelines for redundancy selection in distributed databases
Process with limited communication per unit timeInfluence of soil’s electrical parameters on lightning stroke-current
evolution
Department of Electrical Engineering
The lightning return stroke forms one of the severest natural sources of
electromagnetic inter- ference for systems, both in the air and soil. Several physical
fields govern this complex physical phenomenon, and most of the engineering
applications resort to much simplifications. Several pertinent aspects are somewhat
unclear, and one such important aspect is the influence of soil’s electrical
properties on the stroke current evolution. This is investigated in the present work in a
theoretical framework using the ‘Self-consistent return stroke’ model. The ‘Finite
difference time domain’ (FDTD) method is suitably adopted for the modelling. The
developed FDTD formulation is then used to investigate and ascertain the role of
soil’s electrical properties on the stroke current evolution. For the first time, it is
shown that the soil’s electrical conductivity has some noticeable influence on the
stroke current magnitude.
Characterization and Modelling of Switching Dynamics of SiC MOSFE
Department of Electrical Engineering
SiC MOSFET is a wide bandgap (WBG) power device and commercially
available in the voltage range of 600-1700V. With superior switching, conduction, and
thermal performance, it is in close competition with the state-of-the-art Si IGBTs
in this voltage range.
Fast switching transient of SiC MOSFET reduces switching loss but may induce
prolonged oscillations, spurious turn on, high device stress and EMI related issues,
etc. The nonlinearity of the device characteristics and impact of circuit parasitic
makes the switching transient of SiC MOSFET different from its Si counterpart.
So,modelling the switching dynamics and estimate the switching loss of SiC MOSFET
is essential.
Analytical models are derived through the simplification of complex switching
dynamics. Estimation using this type of model is computationally efficient and can
be easily implemented with freely available programming platforms such as C or
Python. This approach is beneficial at the early stage of the converter design when
switching loss and junction temperature need to be evaluated over several operating
points for many available devices from different vendors. Analytical models require
parameters given in the datasheet.
This talk presents an analytical model to capture the switching dynamics of
SiC MOSFET. A detailed non-linear model of the device along with circuit
parasitic is considered. It results in accurate estimation of transition time, switching loss,
(dv/dt), (di/dt), and transient over-voltage. Also, an analytical model to capture
capacitor-assisted turn-off switching transient will be presented. It helps design the
optimal value of the external snubber capacitor for capacitor-assisted zero voltage
switching (ZVS) converter. In this talk, we will present simple measurement
techniques to determine important circuit parasitic that impacts switching dynamics,
such as the common-source inductance. We will also demonstrate software
developed using Python based on the presented analytical model.
Generalized Source emulation approach on micro-alternator
Department of Electrical Engineering
This talk presents an approach to emulate different power system sources
(different synchronous generators with excitation systems). A testbed consists of a
micro-alternator having an open-loop H-bridge chopper as an exciter has been
developed. To replicate a large machine’s dynamics, a digital time constant regulator
has been incorporated to increase the d-axis transient time constant of the
micro alternator. The mathematical models of standard IEEE excitation systems are usedas
references, and their behaviors have been emulated. The proposed excitation
emulation approach is tested under fault and step change in reference terminal voltage.
The experimental results of a Single Machine infinite bus with these excitation
systems have been compared with simulation for validity. This experimental platform
is essential to test the different control and protection prototypes for validating the
technologies.
Further, a dynamic inversion-based non-linear control approach is presented
to replicate different synchronous generators’ behavior on a micro-alternator. This
approach is validated in simulation under different test cases.
Design Of Active Magnetic Bearing
Department of Electrical Engineering
Bearings are crucial parts in any rotating machine, such as a motor, generator,
or turbine. Majority (> 50%) of the reported faults in electrical machines are related to bearing life and maintenance. These problems are particularly pronounced
in high-speed rotating machines. Active magnetic bearing (AMB) is a contactless
bearing, which supports the rotor of a high-speed electric machine or turbine by controlling the magnetic forces produced by a set of electromagnets positioned around
the rotor. AMBs are increasingly used in high-speed aerospace, industrial, and energy applications. Due to its contact-less and lubrication-free operation, AMBs are
suitable for harsh operating conditions (e.g., extreme temperatures, extreme pressures, corrosive environment) as well as for clean environments (e.g., food processingand pharmaceutical industries). This presentation discusses the operating principle
of AMB, its components, and a system-level overview of AMB. The state-of-the-art
differential-mode current control of AMB and the resulting force-current relationship is explained. The main specifications of an AMB, namely, load capacity and
force slew rate, are discussed. The existing design procedure of AMB is based on
linearized magnetic circuit analysis, which cannot account for the saturation in iron.
An improved design procedure is presented, which accounts for magnetic saturation
and ensures that the force generated is proportional to the control current up to
the desired load capacity. The improved design procedure can achieve both desired
load capacity and linear characteristics while balancing the compactness requirement. The improved design also achieves the desired force slew rate, faster dynamic
response, and enhances the stability of the AMB.
An Algorithm for Fitting Passive Equivalent Circuits for Lumped
Parameter Frequency Dependent Transmission Line models
Department of Electrical Engineering
Accurately fitting rational functions to the frequency response of modal impedances
is very crucial for including frequency dependency in lumped parameter models of
transmission lines. Vector fitting is widely used for fitting rational functions to the
frequency response of modal impedancesand then an R-L equivalent circuit is obtained to get lumped parameter transmission line model.A single-step method based
on the properties of Foster equivalent circuit is proposed in this paper to directly fit
an R-L equivalent circuit to the frequency response of modal impedances. A close
enough fitting is achieved using the proposed method with less number of passive
elements. The proposed line model is validated by studying the switching transients
in 400 kV, 765 kV and 1200 kV transmission lines and comparing the results with
the constant parameter cascaded π-model and the Marti’s Model in EMTP-RV. The
prototype of an experimental scaled-down 230 kV line is developed using amorphous
core inductors and the Clarke’s transformation using 1-φ transformers.
Impact Of Inverter Based Generation On Transmission Protection
Department of Electrical Engineering
The generation mix in the electric power grids around the world is undergoing
a significant change from synchronous machines to inverter based resources (IBR).
As a result, it is essential for the conventional system to adapt to these changes
to bring reliability to the bulk power system(BPS). The increasing penetration of
IBR has affected the conventional protection schemes. While synchronous machines
were capable of providing sufficient amount of fault currents, IBR inverter controls
reduce the fault current levels and short circuit strength. This talk presents a
detailed review on the issues faced by conventional protection schemes in presence
of an IBR.
Switched Capacitor Converters for High Power Applications
Department of Electronic Systems Engineering
The switched capacitor (SC) converters are more compact and have higher
power density when compared with the conventional switched inductor converters.
Due to this, SC converter is being considered for high-power DC-DC applications
like offshore HVDC, electric vehicles and grid integration. A new SC converter topology is presented with minimum capacitor volt-ampere rating and minimum voltage
rating for active switches. The converter is then modified with multiple resonant
inductors to achieve zero-current switching, reducing the switching losses considerably. Two variants of the proposed SC converter topology are further discussed with
a reduced number of resonant inductors.
Do deep networks see the way we do?
Department of Electrical Communication Engineering
Deep neural networks have recently revolutionized computer vision with their
impressive perfor- mance on vision tasks. Their object representations have been
found to match neural representations in the ventral pathway. But do deep neural
networks see the way we do? This is an important question because it will elucidate
the conditions and computations under which perceptual phe- nomena might arise
in neural networks optimized for object classification. Here, we uncover the qualitative similarities and differences between brains and deep networks by comparing
the representations of human perception eliciting objects. The main findings are:(1) Perceptual phenomena like Thatcher effect, Mirror confusion, and Weber’s law
emerge when deep networks are trained for object recognition; (2) Perceptual phenomena like 3D shape processing, surface invariance, and the global advantage are
absent. These results show us when we can consider deep networks good models of vision and how deep networks can be improved.
Deep Learning Methods For Audio EEG Analysis
Department of Electrical Engineering
Brain activations capturing techniques like electroencephalogram (EEG) capture signals not related to the stimuli, which distort the stimulus-response analysis,
and their effect becomes more evident for naturalistic stimuli. To reduce the intersubject redundancies, the EEG responses from multiple subjects can be normalized.Linear methods are the most prominent methods for single-trial analysis. But, they
assume a simplistic linear transfer function. In this talk, we discuss a deep learning framework for audio-EEG data for intra-subject and inter-subject analyses. This
boosts the signals common across the subjects and improves the intra-subject analysis for each subject. Experiments performed on naturalistic speech and music stimuli
listening datasets show that the deep methods obtain better representations than
the linear methods, and that the results are statistically significant. As an extension
work, we compare the linear and deep methods for the task of speech reconstruction
from the EEG recordings.
Decoding Imagined Speech from EEG using Transfer Learning
Department of Electrical Engineering
In this talk, we present a transfer learning based approach for decoding imagined speech from electroencephalogram (EEG). Rather than extracting features separately from individual EEG channels, features are extracted simultaneously from
multiple EEG channels. This helps in capturing the information transfer between
the cortical regions. To alleviate the problem of lack of enough data for training deep
networks, sliding window based data augmentation is performed. Mean phase coherence (MPC) and magnitude-squared coherence (MSC), two popular measures used
in EEG connectivity analysis are used for extracting the features. These features
are compactly arranged, exploiting their symmetry, to obtain a three-dimensional
“image-like” representation. A deep network with ResNet50 as the base model is
used for classifying the imagined prompts. The proposed method is tested on a publicly available EEG dataset recorded during speech imagery. The accuracies obtained
are comparable to the state-of-the-art methods, especially in decoding prompts of
different complexities
Design and Developement of a Headband for Cortical Response
Extraction
Department of Electronic Systems Engineering
Hearing Deficit is the most prevalent chronic neonatal sensory deficit.
Considering adverse repercussions and congenital deafness statistics, early identification and
subsequent intervention significantly save a newborn from lifelong deficiency.
Brainstem Evoked Response Audiometry (BERA) and Otoacoustic Emissions (OAE) are
the existing methods for Neonatal Hearing Screening (NHS). BERA and OAE do
not scan the complete auditory system. Additionally, the lack of clinicians,
expensive equipment, specialized hospitals with audiometry tools, and patient follow-up
are the challenges for the NHS in India. We developed a quick, non-invasive,
cost effective, objective headband to assess the complete auditory pathway. The cortical
auditory evoked response provides important information about the functionality of
subsequent elements of the auditory pathway, making it a comprehensive signature
to evaluate the entire auditory pathway. We obtained a cortical response
(Mismatch Negativity) from n=3 young adults. We aim to replicate a similar headband
for Neonatal Hearing Screening.
Best influential spreaders identification using network global structural
properties
Department of Computational and Data Science
Influential spreaders are the crucial nodes in a complex network that can act
as a controller in case of epidemic spreading or maximizer for information propagation. We have noticed that each individual network holds different connectivity
structures; complete, incomplete, or in-between based on their components and density. These affect the accuracy of existing indexing methods in the identification of
the best influential spreaders. Thus, no single indexing strategy is sufficient for all
varieties of network connectivity structures. This article proposes a new indexing
method Network Global Structure-based Centrality (ngsc), which intelligently combines existing kshell and sum of neighbors’ degree methods with knowledge of the
network’s global structural properties, the giant component, average degree, and
percolation threshold. The experimental results show the proposed method yields a
better spreading performance over a large variety of network connectivity structures
and correlates well with the SIR model used as ground truth.
Modelling Electrical Transport in Biological Tissues
Department of Electronic Systems Engineering
Centre for BioSystems Science and Engineering
Electrical transport in disordered systems such as biological tissues is
characterized by scaling laws which provide a quantifiable description of the system. In
this work, two modes of transport viz. temperature and frequency dependent
electrical transport, are studied using formalin-fixed breast biopsy tissues. Experiments
were performed using an in-house electronic platform integrated with a device that
has on-chip electrodes and a microheater. Temperature dependent direct current
(DC) transport studies were modelled under the realm of general effective medium
theory, with critical temperature (Tc) as a model fit parameter showing higher value
for adjacent normal compared to tumour tissue. Frequency dependent alternating
current (AC) transport studies were modelled using a universal scaling law, with
onset frequency fc as a model fit parameter showing higher values for adjacent normal
samples compared to tumour indicating higher disorder in tumour. We use these
model fit parameters to delineate normal from tumour breast biopsy tissues.
Sensitivity Analysis of ANNs for Fluid Flow Simulation
Department of Computational and Data Sciences
Numerical simulation of convection dominated flows requires numerical schemes
which are high-order accurate and stable. A numerical scheme can be stabilized by
adding the right amount of artificial dissipation to the flux residual. Our ongoing
work focuses on developing a strategy in which the traditional finite element methods
are augmented with Artificial Neural Networks (ANN) to enhance their stability
and efficiency. However, selection of hyper-parameters of the ANN is a non-trivial
task. In this work, we try to address this issue by carrying out a systematic global
sensitivity analysis of ANN hyper-parameters using a technique called Analysis of
Variance (ANOVA) based on four performance metrics. This analysis highlights the
effect of different hyper-parameters on the performance of the ANN as well as helps
us identify the network architectures better suited for our application
Estimation of stochastic responses due to multiple variabilities in
electromagnetic finite element modelling Fleet Missions
Department of Electrical Communication Engineering
A practical electromagnetic system involves several uncertainties like material
properties, surface defects, orientation of radiating systems etc. These deviationswhich are random and distributed all over the domain cannot be taken into account
from the model specifications in a standard FEM simulation. This work formulates
a way to predict the variation of the electromagnetic system response with respect
to the random variation of the input parameters (material and geometric uncertainties) in a full wave 3D edge element FEM, for electromagnetic problems. A
spec- tral stochastic finite element method (SSFEM) is implemented in 3D full wave
edge element based EM problems successfully for multiple material variations. The
method works well for microwave circuit involving waveguides or microstrip lines.
This is then used to formulate geometric variation where the mesh coordinates are
varied as random parameters. This method is found to be faster than the generally
used stochastic collocation and Monte Carlo methods.
Inverse Problems in 3D Full-Wave Electromagnetics
Department of Electrical Communication Engineering
This talk addresses inverse problems specific to the area of electromagnetics,
arising in three different scenarios. The first problem is 3-D quantitative imaging primarily targeted towards bio-medical applications. Two approaches are proposed for
solving the first problem, a multilevel methodology approach and Machine Learning
classification followed by optimization (ML-OPT) approach. The second problem
is in the domain of high-speed circuits and is focused on synthesis of transmission
line physical parameters given the desired electrical parameters like characteristic
impedance and propagation constant. A forward solver is used to train Neural network for several different configurations for analysis and an optimization algorithm
is used for synthesis. The third problem is focused on finding the source of radiation in an electronic system e.g. an automotive ECU, given the measured field at
the antenna in the radiated emissions setup. A method based on Huygens box is
proposed to quantify the radiation from cable and DUT at each frequency. Some
part of the presented work is used via technology-transfer at Simyog Technology Pvt.
Ltd., an IISc incubated startup, to develop a simulation software called Compliance scope which allows the hardware designer to predict the EMI/EMC performance of
electronics modules from an early design stage.
Comparative Analysis of Topological Structures
Department of Computer Science and Automation
Scientific processes often result in scalar fields, whose comprehensive analysis
leads to a better understanding of the underlying phenomena. The feature rich
nature of the scalar fields demands for a summarized representation which has been
addressed by topological structures like persistence diagrams, merge/contour trees,
etc. Many such processes result in a set of fields which may be related temporally,
be part of an ensemble, or unrelated. Thus it also requires methods to compare
them meaningfully.
We develop two global measures, for merge trees, both based on tree edit
distances. One, based on the assumption that they are ordered rooted trees, and
the second, called mted, a metric, assumes unordered rooted trees. Then, we
develop a local comparison measure lmted, also a metric, to compare hierarchical
substructures of scalar fields. We propose dynamic programming algorithms for all
these measures along with intuitive cost models and applications to highlight their
utility.
SPDE-Net: Predict an optimal stabilization parameter for SUPG
technique to solve Singularly Perturbed PDEs
Department of Computational and Data Sciences
Numerical techniques for solving Singularly Perturbed Differential Equations(SPDE)
suffer low accuracy and high numerical instability in presence of interior and
boundary layers. Stabilization techniques are often employed to reduce the spurious
oscillations in the numerical solution. Such techniques are highly dependent on user
chosen stabilization parameter. Here we propose SPDE-Net, a novel neural
network training technique to predict the stabilization parameter. The prediction task
is modeled as a regression problem and is solved using semi supervised learning.
Global and local variants of stabilization parameter τ are demonstrated.
Experiments on a bench-mark case of 1-dimensional convection diffusion equation show a
reasonable performance as compared to the conventional supervised training. This
makes the proposed technique eligible for extension to higher dimensions and other
cases where the analytical formula for stabilization parameter is unknown, therefore
making supervised learning impossible.
Novel p-type AlTiO Gate Oxide based e-mode Operation in AlGaN/GaN HEMTs: 600V Technology with Record High Performance
Department of Electronics Systems Engineering
AlGaN/GaN High Electron Mobility Transistors (HEMTs) are potential candidates for power applications. However, while enhancement-mode (e-mode) is preferred for power applications, AlGaN/GaN HEMTs are depletion mode devices. This work experimentally demonstrates e-mode AlGaN/GaN HEMTs by introducing a novel p-type high-κ AlxTi1−xO based ternary gate oxide. The threshold voltage (Vth) of the GaN HEMTs was found to be tuned by the concentration of Al in Al-Ti-O system, with increasing Al% resulting in a higher positive shift in Vth. Ternary oxide AlTiO was also demonstrated to be superior to regular binary p-oxides, like CuO and NiOx. Using the high-κ and p-type AlTiO, in conjunction with a thinner AlGaN barrier under gate, 600 V e-mode GaN HEMTs are demonstrated with superior ON-state performance (ION ∼ 400 mA/mm and RON = 8.9 Ω-mm) and gate control over channel (ION/IOFF = 107 , SS = 73 mV/dec and gate leakage < 200 nA/mm), beside improved safe operating area reliability
Unique Reliability Phenomena in AlGaN/GaN HEMTs
Department of Electronics Systems Engineering
The Schottky Barrier Diode is a key component in any power electronic circuit or THz/mmW system. A comprehensive TCAD and experimental co-design strategies have been proposed for high power and THz SBD. The critical part of the SBD diode design involves modelling the non-idealities at the Schottky interface and designing physics based process experiments to fix these non-idealities. We have modelled the anode contact interface by accounting for these effects. In addition impact of recessed depth and nature of buffer traps are investigated. The repercussions of field plate design on other performance figures of merit parameters such as diode current collapse, reverse recovery time, reverse current overshoot, and electro-thermal behaviour is investigated. Using the systematic device design approach, we have experimentally demonstrated high power SBD with 15A forward current at 5.5V while having reverse blocking greater than 500V. The design metrics for THz SBD vastly differs from the one used for high power applications. It is imperative to account for the impact of associated device parasitic elements at high-frequency operating conditions. The planer, multi-finger SBD topology looks most promising due to the ease of integration and high cut-off frequencies demonstrated. We present the first report on the design and engineering of multi-finger THz SBD. The study investigates the design metrics of AlN/GaN-based multi-finger, lateral SBD and proposes guidelines to maximize THz operation performance.
Gallium Nitride Hetero-structure based Schottky Barrier Diodes for
Power and RF Applications
Department of Electronics Systems Engineering
AlGaN/GaN High Electron Mobility Transistors (HEMTs) are potential candidates for power applications. However, they suffer from wide range of reliability challenges under different operating conditions. In this work, we explore - (1) dynamic on-resistance (RON) increase post OFF-state stress, (2) gate leakage (IG) current degradation during semi-ON state stress, and (3) OFF-state breakdown physics employing an electro-optic characterization setup and a well-calibrated TCAD simulation setup. Unique critical voltage (VCr) associated with dynamic RON effects was observed and attributed to carrier injection and trapping in Carbon doped GaN buffer governed by channel electric field. Similarly, unique VCr was noticed in semi-ON state beyond which IG degraded permanently. It was attributed to crack/ pit formation in buffer governed by thermo-elastic stress buildup owing to hot electron-buffer trap interaction. OFF-state breakdown in GaN HEMTs was found to be a strong function of surface/buffer trap concentration and device design parameters based on which device design guideline have been established.
Nonlinear Light Generation in Dielectric Metasurfaces
Department of Electronics Systems Engineering
Electro-Thermal Stress Induced Material Reconfiguration & Device
Perturbations in monolayer 2D-TMD Transistors
Department of Electronics Systems Engineering
Device and material reliability of 2-dimensional materials, especially CVD-grown MoS2, has
remained unaddressed since 2011 when the first TMDC transistor was reported. For its potential application in next generation electronics, it is imperative to update our understanding of mechanisms through which MoS2 transistors’ performance degrades under long-term electrical stress. We report,for CVD-grown monolayer MoS2, the very first results on temporal degradation of material and device performance under electrical stress. Both low and high field regimes of operation are explored at different temperatures, gate bias and stress cycles. During low field operation, current is found to saturate after hundreds of seconds of operation with the current decay time constant being a function of temperature and stress cycle. Current saturation after several seconds during low field operation occurs when a thermal equilibrium is established. However, high field operation, especially at low temperature, leads to impact ionization assisted material and device degradation. It is found that high field operation at low temperature results in amorphization of the channel and is verified by device and Kelvin Probe Force Microscopy (KPFM) analyses. In general, a prolonged room temperature operation of CVD-grown MoS2 transistors lead to degraded gate control, higher OFF state current and negative shift in threshold voltage (VT). This is further verified, through micro-Raman and photoluminescence spectroscopy, which suggest that a steady state DC electrical stress leads to the formation of localized low resistance regions in the channel and a subsequent loss of transistor characteristics. Our findings unveil unique mechanism by which CVD MoS2 undergoes material degradation under electrical stress and subsequent breakdown of transistor behaviour. Such an understanding of material and device reliability helps in determining the safe operating regime from device as well as circuit perspective
Transmission Line Outage Identification - A Sequential Testing
Framework
Department of Electrical Communication Engineering
In this work, the problem of identifying transmission line failures in a power
system is investigated. Identifying such failures in the shortest possible time is of
importance due to cascading nature of these failures. With this motivation, a state
estimation based sequential hypothesis testing procedure to localize the failed lines
is proposed. The primary focus is on single line outages as these are more frequently
occurring failures. The state estimation (SE) is performed using conventional power
measurements (SCADA) and synchronized phasor measurement units (PMUs). The
idea is that if there is an outage, then this information is embedded in the state
estimation results, and this information can be utilized to infer the topology of
the power system. The proposed framework involves a Kalman filter based state
estimation followed by a generalized likelihood ratio testing procedure to locate the
failed lines. This work considers both centralized and decentralized state estimation
approaches. In a centralized approach, all the information from various parts of
the power system is available to the system operator. However, in a decentralized
approach, only limited information is considered to reduce the communication and
computational overhead. The proposed algorithms are evaluated on the IEEE 14
and the IEEE 118 bus systems, and results show that all the high risk line failures
were identified quickly. Furthermore, simulation results show that the use of PMUs
enables accurate and quicker detection than conventional power measurements.
Scalable Asynchrony-Tolerant PDE Solver for Multi-GPU System
Department of Computational and Data Science
Data communication and synchronization between processing elements are
likely pose a bottleneck in scalability of partial differential equation(PDE) solvers
on future exascale machines. Recently, asynchrony-tolerant(AT) finite-difference
schemes, which compute accurate solutions while relaxing data movement and
synchronization, were developed using wider stencils in space and/or time to improve
the scalability. As CPU-hardware struggles to keep pace with Moore’s law, GPUs
are significantly contributing towards the compute capability of supercomputers.
However, the GPUs, which possess high throughput, suffer high latency costs. In
this work, we use AT schemes to develop an asynchronous solver for multi-GPUs and
demonstrate their capability to hide the high latency costs. The code, implemented
with CUDA+MPI, solves the compressible fluid-flow equations. Two algorithms
—communication and synchronisation avoiding— were used to develop the solver.
Comparison against a synchronous solver was performed to verify the accuracy and
demonstrate the scalability. On a single node with 8 A100 GPUs a 2X speed-up was
observed. Scalability study on a large number of nodes is currently under progress.
Division of EECS
Electrical, Electronics and Computer Sciences
Indian Institute of Science, Bangalore
Participating Departments
Important Links
© Copyright 2021 All Rights Reserved | Division of EECS