publications

2024

1. Voice Cloning for Low-Resource Languages: Investigating the Prospects for Tamil

   Vishnu Radhakrishnan, Aadharsh Aadhithya, Jayanth Mohan, M Visweswaran, G Jyothish Lal, and B Premjith

   Automatic Speech Recognition and Translation for Low Resource Languages, 2024

   Abs HTML

   With the emergence of artificial intelligence (AI)-powered personalized assistive tools, the surge of futuristic AI agents, and democratization of AI, techniques like voice cloning helps in blurring the line between man and machine. Although there are existing methods for voice synthesis, the task of voice cloning is challenging because the model needs to adapt to an unseen speaker with very less data. Voice cloning is a relatively new task that has not received much attention until recently. While traditional text-to-speech (TTS) systems tries to aid man-machine interaction, voice cloning takes it a step further by enabling to replicate the voice of near or dear ones. However, it is practically difficult to gather large datasets for voice cloning in domestic environments. Apart from the major limitation of data unavailability, designing a compact, mobile, and efficient model for cloning voices with only a few samples of data remains an unaddressed problem. While voice cloning models continue to improve, it remains challenging to incorporate region-specific accents and indigenous low-resource languages into machine-generated audio outputs that accurately differentiate a human voice from a synthesized one. This is primarily because speech synthesis and cloning require a large amount of data, which is not available for low-resource languages. This work examines some preliminary results for voice cloning in Tamil. Despite being in the early stages, our results show promise for the development of a successful voice cloning system for Tamil. We believe that this work will serve as an invitation for the Tamil speech processing community to explore this exciting area of research further. With the potential to revolutionize the field of speech synthesis, voice cloning in Tamil could have significant implications for the development of speech-based applications and assistive technologies. The present work addresses the above-listed problems, specifically for Tamil.

2023

1. Prospects of NFMD for Power System Frequency and Amplitude Estimation

   Aadharsh Aadhithya, Neethu Mohan, S Sachin Kumar, Soman K.P., and Prabhaharan. Poornachandran

   7th IEEE International Conference on Recent Advances and Innovations in Engineering, 2023

   Abs HTML

   Power quality monitoring and parameter estima- tion are essential for the proper functioning of modern power grids. Various techniques have been proposed to estimate the characteristics of power signals and to gain insight into their dynamics. Non-stationary Fourier mode decomposition (NFMD) is a new interpretable time-frequency analysis framework for non-stationary and nonlinear signals. This work investigates the prospects of NFMD in the estimation of power-system characteristics such as fundamental frequency, amplitude and presence of disturbance components. Usage of NFMD as a noise removal technique is also explored in the current work. Results show that NFMD is robust to noisy disturbances and can perform well in noisy environments. The proposed methodology can be used for real-time analysis and interpretation of large volume of data from advanced smart metering techniques such as micro- phasor measurement units in microgrids

2. Exploiting Graph Matrix Duality for Efficient Graph Data Processing

   Aadharsh Aadhithya, Vinith. Rejathalal, Soman K.P., and Prabhaharan Poornachandran

   7th IEEE International Conference on Recent Advances and Innovations in Engineering, 2023

   Abs HTML

   Linear algebra is the backbone of modern applied data analytics. Not only does it provide the theoretical substratum for the development of analytical algorithms it also lends the technological support for extremely efficient implementations. The linear algebra ecosystem consists of remarkably fast hardware support with equally efficient software solutions leveraging the hardware capacity. Graphs are an ubiquitous modelling framework for the representation and analysis of connected data.A flipside from practical perspective is that conventional graph algorithms are challenging as they are hard to parallelize and often make too much use of non-linear data structures which do not have hardware support. In this context, this paper presents an invitation to an alternate way to look at graph processing in terms of algebraic constructs, formulations and technology. The paper presents the theoretical underpinnings of the new view and also some interesting progresses at the frontier level in this direction. Above all, an intended objective is to present a compelling invite to the academia and research community to adopt this approach in their academic and research activities.

3. Optimal Perfect Phylogeny Using ILP and Continuous Approximations

   BE Pranav Kumaar, Aadharsh Aadhithya, S Sachin Kumar, Harishchander Anandaram, and KP Soman

   In International Conference on Advances in Computing and Data Sciences, 2023

   Abs HTML

   The concept of perfect phylogeny is observed to be non-universal, however, in some studies, it is shown that it provides great insights from a biological standpoint. Most natural phylogenies are imperfect but this is debated by the presence of noise in data acting as a disguise, the solution to uncover the underlying perfect phylogeny in the simplest approach is known as Minimum Character Removal (MCR) problem. Another variant of the solution achieved by a change in perspective and computational methodology is known as Maximal Character Compatibility (MCC) problem. Both MCR and MCC problems are solved optimally with reasonable efficiency by using Integer Linear Programming (ILP) technique. A central part of MCC solution involves solving the Max-clique problem of the generated representation allowing numerous clique-solving algorithms to generate various solutions. The comparison between the solution’s optimality and the run-time of these methods substantiates the goal of the different algorithms being applied. The above methods are formulated and implemented using modern programming languages like Python and Matlab. This exploration introduces the problem of perfect phylogeny, its ILP formulations, and its solution along with comparisons between different methods in consideration.

4. Comparitive Study of Continuous Approximations and ILP for Maximal Clique Problems

   Aadharsh Aadhithya, Sachin Kumar, Rejathalal Vinith, Prabaharan Poornachandran, Sujadevi VG, and  K.P

   2023

   Abs

   The maximal clique problem has been a cornerstone for developments in combinatorial optimization. The problem also shows up in innumerable practically relevant scenarios. With rapid advancements in Integer Linear Programming(ILP) solvers, it becomes necessary to have timely benchmark studies against state of the art methods to check if ILP solvers are ready for practical use cases. This work presents a benchmark study between ILP, BK algorithm and few other approximate methods with standard DIMACS instances. This work also aims in providing a friendly introduction to clique based graph problems and one of its application in biological contexts, prediction of missing edges in a protein-protein interaction network. This paper also outlines in supplementary section, simple methods to solve clique problems using excel,python and matlab.

5. DMD-CGR: Dynamic Mode Decomposition-based Novel Features for DNA sequence classification

   Aadharsh Aadhithya, Anirudh Edpuganti, Sachin Kumar, Neethu Mohan, and  others

   In 2023 3rd International Conference on Intelligent Technologies (CONIT), 2023

   Abs HTML

   Accurate classification and analysis of pathogenic organisms are essential for understanding the evolutionary relationships between different viruses, as well as for developing effective treatments and preventive measures. In this paper, we propose the use of machine learning models for pathogen classification, specifically for classifying a novel pathogen at a genus level. We propose a novel feature extraction technique called Dynamic Coherent Features through Progressive Rank Approximations (DCFPRA) for extracting relevant features from genomic sequences. We then evaluate the performance of several machine learning algorithms, including Support Vector Machines (SVM), and K-Nearest Neighbours (KNN), using these features. Our results demonstrate that machine learning models trained on DCFPRA features outperform traditional alignment-based methods for pathogen classification. Furthermore, our proposed method is computationally efficient and can handle large-scale datasets, making it a valuable tool for future research in pathogen classification

6. Finding Network Motifs: A comparative study between ILP and Symmetric Rank-One NMF

   Aadharsh Aadhithya, Sachin Kumar, Vinith Rejathalal, VG Sujadevi, Prabaharan Poornachandran, and KP Soman

   In 2023 3rd International Conference on Intelligent Technologies (CONIT), 2023

   Abs HTML

   This paper compares the performance of two popular methods, an integer linear programming (ILP) based method and a non-negative matrix factorization (NMF) based approximation, for identifying network motifs. Our analysis demonstrates that the ILP-based method is highly accurate but computationally expensive, whereas the NMF-based method is faster and more scalable but less accurate. The choice of method should depend on the specific research question and the size and complexity of the network being analyzed. We also highlight the need for further research to develop more efficient and accurate methods for identifying network motifs, especially for very large and complex networks. Understanding the significance of network motifs and their functions can lead to new therapeutic strategies and improved diagnostics for a wide range of diseases.In addition, we provide an easy-to-use Matlab code implementation for researchers to get started with analyzing their own network motifs. Our code is based on both the ILP-based method, and can be customized to suit the specific needs of the user. This implementation can help facilitate the wider adoption of network motif analysis and enable researchers to gain a deeper understanding of complex systems.

7. An Informal Introduction to Semirings for GraphBLAS using Matlab

   Aadharsh Aadhithya, Anirudh Edpuganti, Vinith Rejathalal, Sachin Kumar, Prabhaharan Poornachandran, and  others

   In 2023 3rd International Conference on Intelligent Technologies (CONIT), 2023

   Abs HTML

   This paper presents an informal introduction to semirings and their application in GraphBLAS using Matlab. GraphBLAS is a powerful and versatile tool for working with graph algorithms and linear algebra, and its use in education is also discussed. The paper simplifies the mathematical complexity typically associated with semirings and provides an accessible entry point for those who are new to the field of GraphBLAS. The use of Matlab as a tool for implementing GraphBLAS is also highlighted, making it user-friendly and easy to understand. Overall, this paper serves as a useful starting point for anyone interested in learning more about GraphBLAS and its potential applications.

8. Learning (With) Distributed Optimization

   Aadharsh Aadhithya, Abinesh S, Akshaya J, Jayanth M, Vishnu Radhakrishnan, Sowmya V, and Soman K. P

   2023

   Abs HTML

   This paper provides an overview of the historical progression of distributed optimization techniques, tracing their development from early duality-based methods pioneered by Dantzig, Wolfe, and Benders in the 1960s to the emergence of the Augmented Lagrangian Alternating Direction Inexact Newton (ALADIN) algorithm. The initial focus on Lagrangian relaxation for convex problems and decomposition strategies led to the refinement of methods like the Alternating Direction Method of Multipliers (ADMM). The resurgence of interest in distributed optimization in the late 2000s, particularly in machine learning and imaging, demonstrated ADMM’s practical efficacy and its unifying potential. This overview also highlights the emergence of the proximal center method and its applications in diverse domains. Furthermore, the paper underscores the distinctive features of ALADIN, which offers convergence guarantees for non-convex scenarios without introducing auxiliary variables, differentiating it from traditional augmentation techniques. In essence, this work encapsulates the historical trajectory of distributed optimization and underscores the promising prospects of ALADIN in addressing non-convex optimization challenges.

9. Harnessing Character-level Dynamics and Word-level Semantics: A New Perspective on Twitter Sentiment Analysis

   Aadharsh Aadhithya, Vishnu Radhakrishnan, Madhav Kishor, Sachin Kumar, and KP Soman

   In 2023 2nd International Conference on Futuristic Technologies (INCOFT), 2023

   Abs HTML

   In recent years, Sentiment Analysis has emerged as an indispensable component of social media analysis. It not only allows understanding public opinion on various issues, but also aids in monitoring real-time trends and sentiments, critical for domains such as brand management, market research, political campaign management, and crisis management. Central to the effectiveness of sentiment analysis is the extraction of meaningful features from the text data, with sentence embeddings being a vital component in capturing semantic relationships and sentence structures. This paper presents a novel approach to Twitter sentiment analysis that exploits sentence embeddings derived from Dynamic Mode Decomposition (DMD). DMD, a dimensionality reduction technique, enables us to capture the underlying dynamics of character transitions in tweets and to construct expressive sentence embeddings. The advantage of DMD lies in its ability to represent complex patterns and dependencies within the tweet text, thereby improving the richness of the derived sentence embeddings. Additionally, we investigate the use of Graph Neural Networks (GNNs) for sentiment classification, harnessing the power of relational modeling in the context of Twitter data. We also delve into word-level analysis with different emotion-based embeddings, assessing their impact on the sentiment classification performance. Our methodology integrates character-level analysis, graph-based modeling, and word-level semantics, providing a comprehensive, multi-layered approach to sentiment analysis on Twitter. Our experiments yield insightful results, highlighting the potential of DMD-based embeddings and the importance of high-quality, context-rich word embeddings in sentiment analysis tasks. These findings contribute to our understanding of sentiment analysis on social media platforms and suggest promising directions for future research in this field.

10. Transfer Learning Approach for Differentiating Parkinson’s Syndromes Using Voice Recordings

    N Sai Satwik Reddy, A Venkata Siva Manoj, V Poorna Muni Sasidhar Reddy, Aadharsh Aadhithya, and V Sowmya

    In International Advanced Computing Conference, 2023

    Abs HTML

    Parkinson syndromes are a group of disorders affecting the elderly population with unsteadiness, slowness of activities, frequent falls, and speech disturbances, which slowly progress. Diagnosis of this group of syndromes is usually purely clinical and could be delayed due to its varied presentations. Parkinson’s syndromes comprise of Idiopathic Parkinson’s disease, Multiple system atrophy (MSA), progressive supranuclear palsy (PSP), and cartico basal ganglionic degeneration. In this work, we provide a comparative analysis of several deep learning models such as ViT, MobileNetV2, DenseNets, ResNets, GoogLeNet, VGGs for the differentiation of Parkinson’s syndromes using prolonged vowel phonations. To address this multi-class classification problem, we employed transfer learning on DL models, by training on a dataset comprising 337 sustained vowels from patients with parkinson’s disease, MSA, PSP, and no parkinson syndromes. Each recording is transformed into a mel-spectrogram for input into the models. Among the models, ResNet152 outperformed the other models, achieving an impressive accuracy of 98.30% in classifying parkinson disorders, offering a promising non-invasive, and cost-effective diagnostic tool for early intervention and treatment planning.

11. Higher Order Dynamic Mode Decomposition for Robust Parameter Estimation in Power Grids

    Pinninti Sai Ravula, Hema Radhika Reddy, Aadharsh Aadhithya, Neethu Mohan, Sachin Kumar, and KP Soman

    In 2023 3rd International Conference on Emerging Frontiers in Electrical and Electronic Technologies (ICEFEET), 2023

    Abs HTML

    The stability, effectiveness, and high-quality operation of the electrical system are maintained through the monitoring of fundamental parameters like frequency and amplitude. The current research provides a data-driven hybrid technique for the extraction of fundamental parameters using the Higher-order dynamic mode decomposition (HODMD) algorithm in smart grid. HODMD’s capacity to account for non-linearities enables it to produce accurate representations of the underlying dynamics, resulting in better model estimation. Thus, HODMD is an appropriate method for power quality analysis due to its capacity to analyze multidimensional data and non-uniformly sampled data. In this work, the potential of HODMD is examined for the estimation of fundamental frequency, amplitude, and the existence of disturbance components, such as harmonics. In the proposed methodology, HODMD modes are computed using the low-rank approximation and then extracted the amplitude and frequency information from the data. The hyper parameters of the proposed methodology are properly tuned to get an accurate estimate of the underlying parameters. The optimistic results on various synthetic and real-time scenarios indicate that the proposed methodology can be utilized to determine the power system frequency and amplitude in general periodic and quasi-periodic dynamics.

2021

1. Implementation of Hack ALU using Quantum Dot Cellular Automata

   Aadharsh Aadhithya, J. Akshaya, K. Ghaayathri Devi, B. Nithesh, G. Jyothish Lal, and K.P. Soman

   2021 Fourth International Conference on Microelectronics, Signals & Systems (ICMSS), 2021

   Abs HTML

   Quantum Dot Cellular Automata (QCA) technology is one of the emerging next-generation nano-scale technologies, to subdue the limitations of existing CMOS technologies. As researchers continue to work hard to find an alternative to CMOS technology, QCA provides a solution for a faster computer with a smaller size and low power consumption. This paper implements an operation-rich “HACK ALU”, capable of performing many operations. Hack AL U is a versatile design that implements 64 op-erations with just 6 control bits along with novel implementations of adders and multiplexers. This ALU has been designed using a bottom-up approach by beginning the design by constructing the basic gates and progressing to the construction of adders, and multiplexers. The scope of the research is to provide an efficient ALU design that can be integrated in the CPU design. The present work has been implemented and tested on the QCA designer software. Experimental evaluation shows that logical and arithmetic operations are consistent in the proposed design.