Kamalesh Karmakar
Kamalesh Karmakar has a total of 30 co-authors. Here are the top co-authors by number of shared publications:
- Sunirmal Khatua
University of Calcutta, Kolkata, India
6 shared publications - Rajib Kumar Das
University of Calcutta
5 shared publications - Arnab Mitra
Techno International New Town
3 shared publications Ana-Maria Ciobotaru
2 shared publications- Anurina Tarafdar
Department of Computer Science and Engineering, University of Calcutta, Kolkata, India
2 shared publications - Kalyan Chakravarthy Kodela
International Technological University, San Jose, USA
2 shared publications
Latest Publications
- 6 October 2026
- Published by Springer Nature
- p. 295-308
An Approach to Detect Fake Profiles in Social Networks Involving an Entropy-Based Naive Bayes Classifier
- 6 October 2026
- Published by Springer Nature
- p. 295-308
An Approach to Detect Fake Profiles in Social Networks Involving an Entropy-Based Naive Bayes Classifier
- Open Access
- 24 August 2026
- Published by Elsevier in SSRN Electronic Journal
Scaling Deep Network Training
Training a modern deep network at scale is no longer a single decision but three coupled ones: which optimizer to trust with millions or billions of parameters, what numerical precision to compute in, and how to spend a fixed compute budget between model size and data. This manuscript develops each decision in turn and then reconnects them. It traces the optimizer lineage from momentum-corrected SGD through Adam, AdamW, and the layer-wise adaptive methods (LARS, LAMB) that make very large batches trainable, develops the mixed-precision training loop and the loss-scaling mechanism that keeps it numerically stable, derives the linear learning-rate scaling rule and its documented limits, and closes with the compute-optimal scaling laws that reframe model size itself as a quantity to be optimized rather than maximized.
- 24 August 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessScaling Deep Network Training
Training a modern deep network at scale is no longer a single decision but three coupled ones: which optimizer to trust with millions or billions of parameters, what numerical precision to compute in, and how to spend a fixed compute budget between model size and data. This manuscript develops each decision in turn and then reconnects them. It traces the optimizer lineage from momentum-corrected SGD through Adam, AdamW, and the layer-wise adaptive methods (LARS, LAMB) that make very large batches trainable, develops the mixed-precision training loop and the loss-scaling mechanism that keeps it numerically stable, derives the linear learning-rate scaling rule and its documented limits, and closes with the compute-optimal scaling laws that reframe model size itself as a quantity to be optimized rather than maximized.
- Open Access
- 19 August 2026
- Published by Elsevier in SSRN Electronic Journal
Geometric Deep Learning for Scientic AI: Graph Neural Networks, AlphaFold, and Neural Operators
Most articial intelligence systems treat all problems the same way: reshape data into rectangular arrays and pass it through mathematical functions. This approach works well for images and text but fails badly for molecular structures, protein shapes, and physical systems where the underlying geometry matters. This paper describes three techniques that build the right geometry into the network from the start. Graph neural networks let atoms talk to nearby atoms through message passing, enabling rapid prediction of molecular properties. AlphaFold predicts protein structures by tracking which amino acids are likely to be close together based on evolution and basic chemistry. Neural operators learn to solve dierential equations by operating directly on functions instead of grid points, replacing hours of traditional computation with near-instant predictions. Each technique solves dierent types of problems eciently, and using them together produces pipelines that outperform traditional scientic computing on its own. Industry use already shows the value: drug screening accelerates 1000Ö, protein structures compute in seconds not months, and uid dynamics simulations run at interactive speed. The main obstacles involve teaching these networks to work on new problems they haven't seen, keeping their reasoning interpretable, and scaling to massive systems. Using specialized hardwareneuromorphic processors for sparse graphs and photonic chips for spectral mathenables practical deployment.
- 19 August 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessGeometric Deep Learning for Scientic AI: Graph Neural Networks, AlphaFold, and Neural Operators
Most articial intelligence systems treat all problems the same way: reshape data into rectangular arrays and pass it through mathematical functions. This approach works well for images and text but fails badly for molecular structures, protein shapes, and physical systems where the underlying geometry matters. This paper describes three techniques that build the right geometry into the network from the start. Graph neural networks let atoms talk to nearby atoms through message passing, enabling rapid prediction of molecular properties. AlphaFold predicts protein structures by tracking which amino acids are likely to be close together based on evolution and basic chemistry. Neural operators learn to solve dierential equations by operating directly on functions instead of grid points, replacing hours of traditional computation with near-instant predictions. Each technique solves dierent types of problems eciently, and using them together produces pipelines that outperform traditional scientic computing on its own. Industry use already shows the value: drug screening accelerates 1000Ö, protein structures compute in seconds not months, and uid dynamics simulations run at interactive speed. The main obstacles involve teaching these networks to work on new problems they haven't seen, keeping their reasoning interpretable, and scaling to massive systems. Using specialized hardwareneuromorphic processors for sparse graphs and photonic chips for spectral mathenables practical deployment.
- 7 May 2026
- Published by Springer Nature
- p. 29-53
Stay-Safer: An IoT-Based On-Demand Monitoring System for Elderly Assistance
- 7 May 2026
- Published by Springer Nature
- p. 29-53
Stay-Safer: An IoT-Based On-Demand Monitoring System for Elderly Assistance
- 23 May 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 1-6
On Tuning the Randomness in the Random Forest Algorithm for Performance Enhancement Towards Fake Profile Detection in Online Social Platforms
The increased popularity of online social networking platforms has positively and negatively impacted users (humans). Among several others, one significant concern is the presence of fake profiles, which may harm users’ experience and trust. To deal with such an issue, we present our research to focus on detecting counterfeit profiles in online social networking platforms. As an instance of such an online platform, we choose Instagram. In our presented research, we have used a supervised Machine Learning Algorithm known as the Random Forest (RF) due to its flexibility and strength towards ensemble learning technique that is especially useful for classification and regression-related tasks. In our proposed approach, we precisely focus on the possible enhancement of the accuracy and efficiency of the fake profile detection technique with enhanced randomness achieved through seed selection. Experimental results confirm that the proposed model achieved 98.71 percent accuracy, an F1-score of 0.96, and an improved Recall value of 97 percent (from 85 percent), which ensures better detection of fake profiles compared to the Baseline Random Forest model.
- 23 May 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 1-6
On Tuning the Randomness in the Random Forest Algorithm for Performance Enhancement Towards Fake Profile Detection in Online Social Platforms
The increased popularity of online social networking platforms has positively and negatively impacted users (humans). Among several others, one significant concern is the presence of fake profiles, which may harm users’ experience and trust. To deal with such an issue, we present our research to focus on detecting counterfeit profiles in online social networking platforms. As an instance of such an online platform, we choose Instagram. In our presented research, we have used a supervised Machine Learning Algorithm known as the Random Forest (RF) due to its flexibility and strength towards ensemble learning technique that is especially useful for classification and regression-related tasks. In our proposed approach, we precisely focus on the possible enhancement of the accuracy and efficiency of the fake profile detection technique with enhanced randomness achieved through seed selection. Experimental results confirm that the proposed model achieved 98.71 percent accuracy, an F1-score of 0.96, and an improved Recall value of 97 percent (from 85 percent), which ensures better detection of fake profiles compared to the Baseline Random Forest model.
- Open Access
- 1 September 2024
- Published by Walter de Gruyter GmbH in Foundations of Computing and Decision Sciences
- Vol. 49 (3) , 201-231
Deep Learning Model for Detection and Severity Analysis of Car Accidents
On-road car accidents are immensely unfortunate but quite common occurrences worldwide. Instant data-centric and informed decisions of crisis management are rarely experienced due to the absence of real-time car accident detection and severity analysis mechanisms. On this background, the current paper presents a deep learning model for car accident detection and analysis of its severity so that the crisis management activities might follow without any delay saving invaluable human lives. The existing works lack in using time-series data, the proper learning model for accurate prediction, and minimizing the time taken in post-accident scenarios for the victims to receive immediate medical help. This paper introduces the Long Short Term Memory (LSTM) model in conjunction with the Gradient Boosted Regression Trees (GBRT) technique for the determination of car accidents with different levels of severity. The proposed model works with the accelerometer and gyroscopic data collected through an application installed in the smartphones of the users inside the car. The LSTM-GBRT hybrid model is proposed to achieve higher accuracy than LSTM which deals with time-variant data. The satisfactory performance of the proposed technique has been reported and the results are extensively investigated in comparison with another hybrid technique such as LSTM with Random Forest (RF) as well. The statistics confirm the superiority of the proposed model over other parallel models in terms of several performance metrics, like Accuracy, Precision, etc.
- 1 September 2024
- Published by Walter de Gruyter GmbH in Foundations of Computing and Decision Sciences
- Vol. 49 (3) , 201-231
Open AccessDeep Learning Model for Detection and Severity Analysis of Car Accidents
On-road car accidents are immensely unfortunate but quite common occurrences worldwide. Instant data-centric and informed decisions of crisis management are rarely experienced due to the absence of real-time car accident detection and severity analysis mechanisms. On this background, the current paper presents a deep learning model for car accident detection and analysis of its severity so that the crisis management activities might follow without any delay saving invaluable human lives. The existing works lack in using time-series data, the proper learning model for accurate prediction, and minimizing the time taken in post-accident scenarios for the victims to receive immediate medical help. This paper introduces the Long Short Term Memory (LSTM) model in conjunction with the Gradient Boosted Regression Trees (GBRT) technique for the determination of car accidents with different levels of severity. The proposed model works with the accelerometer and gyroscopic data collected through an application installed in the smartphones of the users inside the car. The LSTM-GBRT hybrid model is proposed to achieve higher accuracy than LSTM which deals with time-variant data. The satisfactory performance of the proposed technique has been reported and the results are extensively investigated in comparison with another hybrid technique such as LSTM with Random Forest (RF) as well. The statistics confirm the superiority of the proposed model over other parallel models in terms of several performance metrics, like Accuracy, Precision, etc.
- 11 March 2024
- Published by Springer Nature in Journal of Grid Computing
- Vol. 22 (1) , 1-21
Cost-efficient Workflow as a Service using Containers
- 11 March 2024
- Published by Springer Nature in Journal of Grid Computing
- Vol. 22 (1) , 1-21
Cost-efficient Workflow as a Service using Containers
- 4 January 2024
- Published by Springer Nature
- p. 34-49
Scheduling of Containerized Resources for Microservices in Cloud
- 4 January 2024
- Published by Springer Nature
- p. 34-49
Scheduling of Containerized Resources for Microservices in Cloud
- 22 November 2022
- Published by Elsevier in Computers and Electrical Engineering
- Vol. 105
Multi-criteria scheduling of scientific workflows in the Workflow as a Service platform
- 22 November 2022
- Published by Elsevier in Computers and Electrical Engineering
- Vol. 105
Multi-criteria scheduling of scientific workflows in the Workflow as a Service platform
- 17 June 2022
- Published by Elsevier in Journal of Network and Computer Applications
- Vol. 205
Utilization aware and network I/O intensive virtual machine placement policies for cloud data center
- 17 June 2022
- Published by Elsevier in Journal of Network and Computer Applications
- Vol. 205
Utilization aware and network I/O intensive virtual machine placement policies for cloud data center