DAD. Geraldine Bessie Amali

D. Geraldine Bessie Amali

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


D. Geraldine Bessie Amali has a total of 48 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Alternate loss functions and regression models that achieve robustness to outliers by modulating the learning rate

    by Mathew Mithra Noel, Yug D. Oswal, Arindam Banerjee, Venkataraman Muthiah-Nakarajan, Geraldine Bessie Amali D

    Most real-world datasets used for training supervised learning models are contaminated with noisy data and outliers that complicate the model parameter learning process. This paper proposes a new approach for achieving robustness where the learning rate is modulated by a factor that is sensitive to outliers. In this approach a reduction of the learning rate is shown to be achieved by using alternate loss functions that are infinitely differentiable, strictly convex or quasiconvex and more closely approximate the absolute error than Huber and log-cosh losses. A comparison of the performance of regression models trained with different loss functions on a wide variety of benchmarks and datasets is presented to demonstrate the superior performance of the Square Root Loss (SRL) and Smooth Mean Absolute Error (SMAE) losses proposed in this paper. Two new robust linear regression models are presented. Highly vectorized robust parameter update formulae that take advantage of modern GPUs for both stochastic and batch gradient descent are presented. Finally, a new robust gradient descent parameter update formula for nonlinear regression models is presented.

  • Open Access

    Alternate Loss Functions for Classification and Robust Regression Can Improve the Accuracy of Artificial Neural Networks

    by Mathew Mithra Noel, Arindam Banerjee, Yug D. Oswal, Geraldine Bessie Amali D, Venkataraman Muthiah-Nakarajan

    All machine learning algorithms use a loss, cost, utility, or reward function to encode the learning objective and oversee the learning process. This paper shows that the training speed, final accuracy and robustness to outliers can significantly depend on the loss function used to train neural networks. In particular derivative of the loss with respect to the output can be significantly different with different loss functions leading to significantly different performance after gradient-descent based training. This paper explores the possible advantages of using new loss functions that penalize errors differently compared to the popular Cross-entropy and Mean squared loss on 7 benchmarks. Two new loss functions for classification are proposed and switching loss functions during training is shown to significantly speedup training. Two new infinitely differentiable loss function that more closely approximate the absolute error loss compared to the Huber and Log-Cosh losses used for robust regression are proposed and shown to significantly outperform Squared error, Huber, and Log-Cosh losses on datasets contaminated with outliers.

  • Open Access

    A new framework with convoluted oscillatory neural network for efficient object-based land use and land cover classification on remote sensing images

    by Chirag Jitendra Chandnani, Shlok Chetan Kulkarni, Geraldine Bessie Amali D, Rohini Selvaraj

    Rigorous urbanization leads to unprecedented climate change. Pune area in India has witnessed recent flash floods and landslides due to unplanned rapid urbanization. It, therefore, becomes vital to manage and analyse man-made impact on the environment through effective land use land cover classification (LULC). Accurate LULC classification allows for better planning and effective allocation of resources in urban development. Remote sensing images provide surface reflectance data that are used for accurate mapping and monitoring of land cover. Convolution neural networks (CNN) trained with Relu are conventionally used in classifying different land types. However, every neuron has a single hyperplane decision boundary which restricts the model's capability to generalize. Oscillatory activation functions with their periodic nature have demonstrated that a single neuron can have multiple hyperplanes in the decision boundary which helps in better generalization and accuracy. This study proposes a novel framework with convoluted oscillatory neural networks (CONN) that synergistically combines the periodic, non-monotonic nature of oscillatory activation functions with the deep convoluted architecture of CNNs to accurately map LULC. Results carried out on LANDSAT-8 surface reflectance images for the Pune area indicate that CONN with Decaying Sine Unit achieved an overall train accuracy of 99.999%, test accuracy of 95.979% and outperforms conventional CNN models in precision, recall and User's Accuracy. A thorough ablation study was conducted with various subsets of the feature set to test the performance of the selected models in the absence of data.


  • Open Access

    Quantum-Enhanced Algorithmic Fairness and the Advancement of AI Integrity and Responsibility

    by Akhil Chintalapati, Khashbat Enkhbat, Ramanathan Annamalai, Geraldine Bessie Amali, Fatih Ozaydin, Mathew Mithra Noel

    In the evolving digital landscape, the pervasive influence of artificial intelligence (AI) on social media platforms reveals a compelling paradox: the capability to provide personalized experiences juxtaposed with inherent biases reminiscent of human imperfections. Such biases prompt rigorous contemplation on matters of fairness, equity, and societal ramifications, and penetrate the foundational essence of AI. Within this intricate context, the present work ventures into novel domains by examining the potential of quantum computing as a viable remedy for bias in artificial intelligence. The conceptual framework of the quantum sentinel is presented—an innovative approach that employs quantum principles for the detection and scrutiny of biases in AI algorithms. Furthermore, the study poses and investigates the question of whether the integration of advanced quantum computing to address AI bias is seen as an excessive measure or a requisite advancement commensurate with the intricacy of the issue. By intertwining quantum mechanics, AI bias, and the philosophical considerations they induce, this research fosters a discourse on the journey toward ethical AI, thus establishing a foundation for an ethically conscious and balanced digital environment. We also show that the quantum Zeno effect can protect SVM hyperplanes from bias through targeted simulations.


  • Open Access

    A Physics-Based Hyper Parameter Optimized Federated Multi-Layered Deep Learning Model for Intrusion Detection in IoT Networks

    by Chirag Jitendra Chandnani, Vedik Agarwal, Shlok Chetan Kulkarni, Aditya Aren, D. Geraldine Bessie Amali, Kathiravan Srinivasan

    The Internet of Things (IoT) is reshaping our lives with its omnipresence. The sudden uptick in the ubiquitous nature of IoT devices ranging from fitness watches to aircraft has led to a surge of cyber-attacks. Artificial Intelligence powered Intrusion Detection Systems (IDS) are being used recently to combat this increasing surge of attacks in the IoT environment. However, existing solutions lack optimization for training in distributed decentralized environments. A popular solution for training a model in a decentralized environment is Federated Learning. Multiple client models collaboratively train a global model while keeping the individual client’s data decentralized and private. This, however, suffers from poor generalization of the individual client data. This work proposes a new Federated Multi-Layered Deep-Learning (Fed-MLDL) model that employs physics-based hyperparameter optimization (HPO) technique FedRIME in a distributed federated learning environment for intrusion detection on the CICIoT23, CICIoT22, ToN_IoT, Edge_IIoT and, IoT-23 datasets. FedRIME ensures good generalization for all clients’ data by finetuning the model’s hyperparameters according to each client.The experimental results indicate that the Fed-MLDL with Fed-RIME optimization exhibits the highest accuracy for independent and identically distributed datasets with the scores being 99.2% with CICIoT23, 98.1% with CICIoT22, 98.2% with ToN_IoT, 98.5% with Edge_IIoTset and, 98.6% with IoT-23 dataset respectively. Further, the proposed Fed-MLDL with Fed-RIME optimization has demonstrated a significant improvement in the speed of convergence, stability, and client specific customization in federated learning. The study provides a comprehensive comparison with the most recent physics based HPO techniques. This study observes that coupling a Deep-Learning model with HPO techniques results in a much faster convergence requiring only 10-15 communication rounds. The proposed Fed-MLDL with Fed-RIME optimization outperforms existing state-of-the-art models on the CIC-IoT23 dataset.


  • Open Access

    A New Texture Aware—Seed Demand Enhanced Simple Non-Iterative Clustering (ESNIC) Segmentation Algorithm for Efficient Land Use and Land Cover Mapping on Remote Sensing Images

    by Rohini Selvaraj, D. Geraldine Bessie Amali

    Change detection in Land Use and Land Cover (LULC) on remote sensing images is essential for urban planning, disaster risk management, climate change monitoring, and biodiversity conservation. Precise detection of these changes is heavily impacted by the classification accuracy of the LULC types which can be improved significantly by addressing the misclassification errors arising due to similar spectral LULC types and overlapping LULC regions. This paper proposes a texture-aware and seed-demanding Enhanced Simple Non-Iterative Clustering (ESNIC) segmentation algorithm and Boundary-Specific Two-Level (BSTL) classification approach that reduces misclassification rates due to similar spectral signatures and minimizes computational redundancy. Incorporating texture features extracted through the Gray-Level Co-occurrence Matrix along with spectral information in the proposed ESNIC segmentation algorithm improves the ability to distinguish between different LULC types that share the same spectral value. The seed demanding ESNIC segmentation approach seeds are strategically placed based on the content adaptation approach rather than being uniformly distributed throughout the image which reduces segmentation time, providing a substantial advantage for large-scale land cover mapping. A BSTL classification approach that synergistically combines the Support Vector Machine’s ability to effectively handle high dimensional data with the k-Nearest Neighbor’s ability to handle irregular data is used. This study is assessed in terms of Overall Accuracy(OA), Producer Accuracy, User Accuracy, kappa coefficients (K), Root Mean Square Errors (RMSE), and F1 scores. Results indicate that the proposed ESNIC-BSTL (OA = 97.18%, K= 0.96 and RMSE = 0.1311) approach provides better accuracy than SNIC-SVM (94.42%, 0.92, and 0.1422) and SNIC- BSTL (95.78%, 0.94 and 0.1362).


  • Biologically inspired oscillating activation functions can bridge the performance gap between biological and artificial neurons

    by Mathew Mithra Noel, Shubham Bharadwaj, Venkataraman Muthiah-Nakarajan, Praneet Dutta, Geraldine Bessie Amali D.


  • Open Access

    Quantum-Enhanced Algorithmic Fairness and the Advancement of AI Integrity and Responsibility

    by Akhil Chintalapati, Khashbat Enkhbat, Ramanathan Annamalai, Geraldine Bessie Amali, Fatih Ozaydin, Mathew Mithra Noel

    In the evolving digital landscape, the pervasive influence of Artificial Intelligence (AI) on social media platforms reveals a compelling paradox: the capability to provide personalized experiences juxtaposed with inherent biases reminiscent of human imperfections. Such biases prompt rigorous contemplation on matters of fairness, equity, and societal ramifications, and penetrate the foundational essence of AI. Within this intricate context, the research ventures into novel domains by examining the potential of quantum computing as a viable remedy for bias in artificial intelligence. The conceptual framework of the Quantum Sentinel is presented - an innovative approach that employs quantum principles for the detection and scrutiny of biases in AI algorithms. Furthermore, the study poses and investigates the question of whether the integration of advanced quantum computing to address AI bias is seen as an excessive measure or a requisite advancement commensurate with the intricacy of the issue. By intertwining quantum mechanics, AI bias, and the philosophical considerations they induce, this research fosters a discourse on the journey toward ethical AI, thus establishing a foundation for an ethically conscious and balanced digital environment.

  • Accurate classification of land use and land cover using a boundary-specific two-level learning approach augmented with auxiliary features in Google Earth Engine

    by Rohini Selvaraj, Geraldine Bessie Amali D

    Land use land cover (LULC) classification using remote sensing images is a valuable resource in various fields such as climate change, urban development, and land degradation monitoring. The city of Madurai in India is known for its diverse geographical elements and rich heritage, which includes the cultural sport of “Jallikattu”: whose main competitor, the zebusare deeply affected by the conversion of their waterbodies and pastures into concrete jungles. Hence, monitoring land degradation is vital in preserving the geography and cultural heritage of the study area, Madurai. The “Landsat 8 Operational Land Imager tier_2 collection_2 Level_2 Surface Reflectance” image was taken for this study. The LULC classification is performed based on the following classes: forest, agriculture, urban, water bodies, uncultivated land, and bare land. The objective of the study is to incorporate auxiliary features to spectral and textural features along with a simple non-iterative clustering (SNIC) segmentation algorithm and implement a boundary-specific two-level learning approach based on support vector machines (SVM) and k nearest neighbors (kNN) classification algorithms. The overall accuracy (OA) of 95.78% and 0 .94 Kappa score (K) were obtained using a boundary-specific two-level model augmented with auxiliary feature and SNIC algorithm in comparison to PB, OB, and OBS, which achieve OA (K) of 81% (0.76), 91% (0.89), and 94.42% (0.92), respectively. The results demonstrate a notable enhancement in overall classification accuracy when augmenting the features and refining classification decisions using a boundary-specific two-level learning approach.