Adaptive and Parallel Split Federated Learning in Vehicular Edge Computing

Xianke Qiang, Zheng Chang, Yun Hu, Lei Liu, Timo Hamalainen·May 29, 2024

Summary

The paper proposes Adaptive Split Federated Learning for Vehicular Edge Computing (ASFV), a novel approach that combines split learning and federated learning to address challenges in vehicular edge computing. ASFV adapts to vehicle heterogeneity, mitigates privacy risks, and reduces latency on non-IID data. Key contributions include a vehicle selection algorithm, a time delay minimization function, and a decomposition method for handling network dynamics and mobility. The scheme outperforms existing methods in terms of efficiency and privacy, with simulations using open datasets demonstrating improved performance over FL and SFL, especially in terms of accuracy, time delays, and energy consumption. The paper also presents mathematical bounds on error terms and analyzes the convergence of the learning process in the context of vehicle networks.

Key findings

9

Paper digest

What problem does the paper attempt to solve? Is this a new problem?

The paper aims to address the challenge of training complete and large models on resource-constrained vehicles in the context of vehicular edge computing . This problem is not entirely new, but the paper proposes an innovative solution called Adaptive Split Federated Learning for Vehicular Edge Computing Systems (ASFV) to optimize the performance of federated learning in vehicular networks . The ASFV solution combines adaptive split federated training, vehicle selection, and resource allocation to improve time-energy efficiency and learning performance over heterogeneous devices .


What scientific hypothesis does this paper seek to validate?

This paper aims to validate a novel low-latency and low-energy Adaptive Split Federated Training (ASFV) hypothesis by introducing an adaptive split federated training approach that combines vehicle selection and resource allocation. The study conducts a thorough theoretical analysis of the training delay and energy consumption of the proposed ASFV .


What new ideas, methods, or models does the paper propose? What are the characteristics and advantages compared to previous methods?

The paper proposes several novel ideas, methods, and models in the field of vehicular edge computing and federated learning:

  • Adaptive Split Federated Learning for Vehicular Edge Computing Systems (ASFV): The paper introduces an innovative ASFV approach that combines adaptive split federated training with vehicle selection and resource allocation . This approach aims to optimize time-energy efficiency and learning performance for heterogeneous devices in vehicular edge computing systems.
  • Vehicle Selection Algorithm: The paper presents a vehicle selection algorithm based on vehicle speed and EC communication range, contributing to minimizing time delays and energy consumption .
  • Multi-Objective Function Formulation: The paper formulates a multi-objective function that considers resource management, vehicle heterogeneity, channel instability, and model splitting strategy to address the challenges in vehicular edge federated learning .
  • Decomposition of Problem: Due to the complexity of the formulated problem, the paper decomposes it into three subproblems and iteratively solves them using methods like KKT, SCA, and Lagrange multiplier Method to approximate optimal solutions .
  • Resource Management Algorithm: The paper proposes a resource management algorithm that takes into account vehicle heterogeneity, network dynamics, and vehicle mobility to optimize the performance of the ASFV solution over wireless networks .

These proposed ideas and methods aim to enhance the efficiency, performance, and adaptability of federated learning in vehicular edge computing systems, addressing challenges such as resource constraints, dynamic communication environments, and heterogeneous data distributions among vehicles. The proposed Adaptive Split Federated Learning for Vehicular Edge Computing Systems (ASFV) introduces several key characteristics and advantages compared to previous methods outlined in the paper :

  • Adaptive Split Federated Training: ASFV combines adaptive split federated training with vehicle selection and resource allocation, optimizing time-energy efficiency and learning performance for heterogeneous devices in vehicular edge computing systems .
  • Vehicle Selection Algorithm: ASFV incorporates a vehicle selection algorithm based on vehicle speed and EC communication range, contributing to minimizing time delays and energy consumption .
  • Resource Management Algorithm: The proposed solution includes a resource management algorithm that considers vehicle heterogeneity, network dynamics, and vehicle mobility to optimize performance over wireless networks .
  • Multi-Objective Function Formulation: ASFV formulates a multi-objective function that combines resource management, vehicle heterogeneity, channel instability, and model splitting strategy to address challenges in vehicular edge federated learning .
  • Decomposition of Problem: The complex problem is decomposed into three subproblems and solved iteratively using methods like KKT, SCA, and Lagrange multiplier Method to approximate optimal solutions .
  • Performance Evaluation: Extensive simulations using various datasets demonstrate the superiority of ASFV in terms of time-energy efficiency and learning performance for vehicular edge computing systems, showcasing its effectiveness over existing schemes .
  • Comparison with Previous Methods: ASFV exhibits accuracy closest to Sequential Learning (SL) with significantly faster convergence speed compared to traditional Federated Learning (FL) and Split Federated Learning (SFL) methods. It achieves notably higher accuracy than other parallel FL and SFL approaches, although slightly lower than sequential SL .

These characteristics and advantages highlight the innovative nature of ASFV in addressing the challenges of vehicular edge computing systems, offering improved efficiency, performance, and adaptability compared to traditional and existing methods in the field.


Do any related researches exist? Who are the noteworthy researchers on this topic in this field?What is the key to the solution mentioned in the paper?

Several related research papers exist in the field of federated learning and vehicular edge computing. Noteworthy researchers in this area include G. Poirot, P. Vepakomma, K. Chang, J. Kalpathy-Cramer, R. Gupta, R. Raskar, W. Wu, M. Li, K. Qu, C. Zhou, X. Shen, W. Zhuang, X. Li, W. Shi, R. Chen, L. Li, K. Xue, C. Zhang, M. Pan, Y. Fang, L. Zhu, F. R. Yu, T. Tang, Z. Zhou, E. Li, L. Zeng, K. Luo, J. Zhang, S. Song, K. B. Letaief, S. Luo, X. Chen, Q. Wu, Z. Zhou, S. Yu, Y. Fu, H. Guo, X. Yang, Y. Ding, V. Chandra, Y. Lin, and many others .

The key to the solution mentioned in the paper involves various strategies such as weight quantization, gradient quantization, joint optimization of local accuracy, transmit power, data rate, and devices' computing capacities, as well as communication-efficient federated learning frameworks with partial model aggregation algorithms and compression strategies. Additionally, split learning (SL) and split federated learning (SFL) have been proposed to address challenges in communication efficiency and training latency, with approaches like Cluster-based Parallel SL (CPSL) and Hybrid Split and Federated Learning (HSFL) being developed to optimize model training mechanisms .


How were the experiments in the paper designed?

The experiments in the paper were designed as follows:

  • The paper proposed a novel low-latency and low-energy Adaptive Split Federated Vehicle (ASFV) training by combining vehicle selection and resource allocation, with a theoretical analysis of training delay and energy consumption .
  • A vehicle selection algorithm based on vehicle speed and EC communication range was formulated, along with a time delay minimization multi-objective function considering vehicle heterogeneity, channel instability, and model splitting strategy .
  • The formulated multi-objective problem was decomposed into three subproblems: online adaptive cut layer selection, transmission power assignment, and wireless resource allocation, which were solved iteratively using BCD, KKT, SCA, and Lagrange multiplier methods .
  • The performance of the proposed solution was evaluated through extensive simulations using various open datasets, demonstrating superior time-energy efficiency and learning performance for ASFV over heterogeneous devices compared to existing schemes .

What is the dataset used for quantitative evaluation? Is the code open source?

The dataset used for quantitative evaluation in the research is a training set with 50,000 samples for model training and a test set with 10,000 samples for performance evaluation . The code for the research is not explicitly mentioned to be open source in the provided context.


Do the experiments and results in the paper provide good support for the scientific hypotheses that need to be verified? Please analyze.

The experiments and results presented in the paper provide substantial support for the scientific hypotheses that needed verification. The paper extensively explores the challenges and advancements in Federated Learning (FL) within Vehicular Edge Computing (VEC) systems, addressing critical issues such as high heterogeneity among vehicles/clients involved in training and the protection of user privacy . The experiments conducted in the paper demonstrate the application of Split Learning (SL) and Federated Learning (FL) in VEC systems, showcasing the potential benefits and challenges associated with these approaches . Additionally, the paper discusses the impact of mobility, communication delays, energy consumption, and resource allocation on the performance of FL and SL algorithms in VEC systems, providing valuable insights into the practical implications of these technologies .

Moreover, the paper introduces innovative approaches such as Split Learning-based Intrusion Detection Systems (IDS) and SplitFed learning with mobility methods to enhance model training efficiency and security in VEC systems . These experiments contribute to the validation of the scientific hypotheses by demonstrating the effectiveness of SL and FL in addressing security concerns and optimizing training times in VEC environments . The results presented in the paper offer a comprehensive analysis of the performance metrics, convergence speeds, and energy consumption associated with FL and SL algorithms in VEC systems, supporting the scientific hypotheses and providing valuable insights for future research in this domain .


What are the contributions of this paper?

The paper makes several contributions in the field of vehicular edge computing and federated learning:

  • Proposing a Split Learning-based IDS (SplitLearn) for ITS infrastructures to address security concerns, outperforming Federated Learning (FedLearn) and Transfer Learning (TransLearn) .
  • Introducing SplitFed learning with a mobility method to minimize model training time and a migration method for the ML model when vehicles move between different VECs .
  • Addressing challenges in Split Learning (SL)-assisted and Split Federated Learning (SFL)-assisted VEC, highlighting the need for detailed analysis on model migration, long SL serial delay, and device allocation for FL and SL .

What work can be continued in depth?

Further research can be conducted to delve deeper into the exploration of Split Learning (SL) as a collaborative learning framework, especially within the context of vehicular networks. While SL has shown promising results in addressing security concerns and outperforming other learning solutions like Federated Learning (FL) and Transfer Learning (TransLearn) , there is still room for in-depth investigation into its application and optimization in vehicular edge computing systems. Specifically, research can focus on enhancing the efficiency and effectiveness of Split Learning in addressing the unique challenges and requirements of vehicular networks, such as optimizing model splitting for supporting a large number of vehicles, considering vehicle heterogeneity, network dynamics, and vehicle mobility . By further exploring adaptive and parallel Split Federated Learning solutions, researchers can aim to optimize performance over wireless networks and improve the overall learning process in vehicular edge computing systems .

Tables

1

Introduction
Background
Evolution of vehicular edge computing
Challenges in FL and SFL for vehicular networks
Objective
To address vehicle heterogeneity, privacy, and latency in VEC
Improve efficiency and privacy compared to FL and SFL
Methodology
Adaptive Federated Split Learning Architecture
1.1 Vehicle Selection Algorithm
Criteria for selecting participating vehicles
Balancing resource diversity and network dynamics
1.2 Time Delay Minimization Function
Formulation and optimization of delay reduction
Handling network dynamics and mobility
Data Processing and Split Learning
2.1 Data Collection
Federated data collection from vehicles
Handling non-IID data distribution
2.2 Data Preprocessing
Data preprocessing techniques for edge devices
Privacy preservation through local computations
Decomposition Method
Handling network dynamics and mobility in model updates
Scalability for large vehicular networks
Performance Evaluation
Simulation Setup and Datasets
Open datasets used for simulations
Comparison with FL and SFL baselines
3.1 Performance Metrics
Accuracy comparison
Time delays analysis
Energy consumption analysis
Error Bounds and Convergence Analysis
3.2 Mathematical Error Bounds
Derivation and analysis of convergence rates
Impact of vehicle mobility on convergence
3.3 Convergence Guarantees
Conditions for convergence in vehicular networks
Effect of adaptive learning on convergence
Conclusion
Summary of key findings
Significance of ASFV in vehicular edge computing
Future research directions
References
List of cited literature and contributions
Basic info
papers
machine learning
artificial intelligence
networking and internet architecture
Advanced features
Insights
How does ASFV address the challenges in vehicular edge computing, specifically in terms of vehicle heterogeneity and privacy risks?
What are the key contributions of the paper in terms of algorithms and methods for improving performance in ASFV?
What is the primary focus of the Adaptive Split Federated Learning for Vehicular Edge Computing (ASFV) proposed in the paper?
How does ASFV compare to FL and SFL in terms of accuracy, time delays, and energy consumption, as demonstrated through simulations with open datasets?

Adaptive and Parallel Split Federated Learning in Vehicular Edge Computing

Xianke Qiang, Zheng Chang, Yun Hu, Lei Liu, Timo Hamalainen·May 29, 2024

Summary

The paper proposes Adaptive Split Federated Learning for Vehicular Edge Computing (ASFV), a novel approach that combines split learning and federated learning to address challenges in vehicular edge computing. ASFV adapts to vehicle heterogeneity, mitigates privacy risks, and reduces latency on non-IID data. Key contributions include a vehicle selection algorithm, a time delay minimization function, and a decomposition method for handling network dynamics and mobility. The scheme outperforms existing methods in terms of efficiency and privacy, with simulations using open datasets demonstrating improved performance over FL and SFL, especially in terms of accuracy, time delays, and energy consumption. The paper also presents mathematical bounds on error terms and analyzes the convergence of the learning process in the context of vehicle networks.
Mind map
Effect of adaptive learning on convergence
Conditions for convergence in vehicular networks
Impact of vehicle mobility on convergence
Derivation and analysis of convergence rates
Energy consumption analysis
Time delays analysis
Accuracy comparison
Privacy preservation through local computations
Data preprocessing techniques for edge devices
Handling non-IID data distribution
Federated data collection from vehicles
Handling network dynamics and mobility
Formulation and optimization of delay reduction
Balancing resource diversity and network dynamics
Criteria for selecting participating vehicles
3.3 Convergence Guarantees
3.2 Mathematical Error Bounds
3.1 Performance Metrics
Scalability for large vehicular networks
Handling network dynamics and mobility in model updates
2.2 Data Preprocessing
2.1 Data Collection
1.2 Time Delay Minimization Function
1.1 Vehicle Selection Algorithm
Improve efficiency and privacy compared to FL and SFL
To address vehicle heterogeneity, privacy, and latency in VEC
Challenges in FL and SFL for vehicular networks
Evolution of vehicular edge computing
List of cited literature and contributions
Future research directions
Significance of ASFV in vehicular edge computing
Summary of key findings
Error Bounds and Convergence Analysis
Simulation Setup and Datasets
Decomposition Method
Data Processing and Split Learning
Adaptive Federated Split Learning Architecture
Objective
Background
References
Conclusion
Performance Evaluation
Methodology
Introduction
Outline
Introduction
Background
Evolution of vehicular edge computing
Challenges in FL and SFL for vehicular networks
Objective
To address vehicle heterogeneity, privacy, and latency in VEC
Improve efficiency and privacy compared to FL and SFL
Methodology
Adaptive Federated Split Learning Architecture
1.1 Vehicle Selection Algorithm
Criteria for selecting participating vehicles
Balancing resource diversity and network dynamics
1.2 Time Delay Minimization Function
Formulation and optimization of delay reduction
Handling network dynamics and mobility
Data Processing and Split Learning
2.1 Data Collection
Federated data collection from vehicles
Handling non-IID data distribution
2.2 Data Preprocessing
Data preprocessing techniques for edge devices
Privacy preservation through local computations
Decomposition Method
Handling network dynamics and mobility in model updates
Scalability for large vehicular networks
Performance Evaluation
Simulation Setup and Datasets
Open datasets used for simulations
Comparison with FL and SFL baselines
3.1 Performance Metrics
Accuracy comparison
Time delays analysis
Energy consumption analysis
Error Bounds and Convergence Analysis
3.2 Mathematical Error Bounds
Derivation and analysis of convergence rates
Impact of vehicle mobility on convergence
3.3 Convergence Guarantees
Conditions for convergence in vehicular networks
Effect of adaptive learning on convergence
Conclusion
Summary of key findings
Significance of ASFV in vehicular edge computing
Future research directions
References
List of cited literature and contributions
Key findings
9

Paper digest

What problem does the paper attempt to solve? Is this a new problem?

The paper aims to address the challenge of training complete and large models on resource-constrained vehicles in the context of vehicular edge computing . This problem is not entirely new, but the paper proposes an innovative solution called Adaptive Split Federated Learning for Vehicular Edge Computing Systems (ASFV) to optimize the performance of federated learning in vehicular networks . The ASFV solution combines adaptive split federated training, vehicle selection, and resource allocation to improve time-energy efficiency and learning performance over heterogeneous devices .


What scientific hypothesis does this paper seek to validate?

This paper aims to validate a novel low-latency and low-energy Adaptive Split Federated Training (ASFV) hypothesis by introducing an adaptive split federated training approach that combines vehicle selection and resource allocation. The study conducts a thorough theoretical analysis of the training delay and energy consumption of the proposed ASFV .


What new ideas, methods, or models does the paper propose? What are the characteristics and advantages compared to previous methods?

The paper proposes several novel ideas, methods, and models in the field of vehicular edge computing and federated learning:

  • Adaptive Split Federated Learning for Vehicular Edge Computing Systems (ASFV): The paper introduces an innovative ASFV approach that combines adaptive split federated training with vehicle selection and resource allocation . This approach aims to optimize time-energy efficiency and learning performance for heterogeneous devices in vehicular edge computing systems.
  • Vehicle Selection Algorithm: The paper presents a vehicle selection algorithm based on vehicle speed and EC communication range, contributing to minimizing time delays and energy consumption .
  • Multi-Objective Function Formulation: The paper formulates a multi-objective function that considers resource management, vehicle heterogeneity, channel instability, and model splitting strategy to address the challenges in vehicular edge federated learning .
  • Decomposition of Problem: Due to the complexity of the formulated problem, the paper decomposes it into three subproblems and iteratively solves them using methods like KKT, SCA, and Lagrange multiplier Method to approximate optimal solutions .
  • Resource Management Algorithm: The paper proposes a resource management algorithm that takes into account vehicle heterogeneity, network dynamics, and vehicle mobility to optimize the performance of the ASFV solution over wireless networks .

These proposed ideas and methods aim to enhance the efficiency, performance, and adaptability of federated learning in vehicular edge computing systems, addressing challenges such as resource constraints, dynamic communication environments, and heterogeneous data distributions among vehicles. The proposed Adaptive Split Federated Learning for Vehicular Edge Computing Systems (ASFV) introduces several key characteristics and advantages compared to previous methods outlined in the paper :

  • Adaptive Split Federated Training: ASFV combines adaptive split federated training with vehicle selection and resource allocation, optimizing time-energy efficiency and learning performance for heterogeneous devices in vehicular edge computing systems .
  • Vehicle Selection Algorithm: ASFV incorporates a vehicle selection algorithm based on vehicle speed and EC communication range, contributing to minimizing time delays and energy consumption .
  • Resource Management Algorithm: The proposed solution includes a resource management algorithm that considers vehicle heterogeneity, network dynamics, and vehicle mobility to optimize performance over wireless networks .
  • Multi-Objective Function Formulation: ASFV formulates a multi-objective function that combines resource management, vehicle heterogeneity, channel instability, and model splitting strategy to address challenges in vehicular edge federated learning .
  • Decomposition of Problem: The complex problem is decomposed into three subproblems and solved iteratively using methods like KKT, SCA, and Lagrange multiplier Method to approximate optimal solutions .
  • Performance Evaluation: Extensive simulations using various datasets demonstrate the superiority of ASFV in terms of time-energy efficiency and learning performance for vehicular edge computing systems, showcasing its effectiveness over existing schemes .
  • Comparison with Previous Methods: ASFV exhibits accuracy closest to Sequential Learning (SL) with significantly faster convergence speed compared to traditional Federated Learning (FL) and Split Federated Learning (SFL) methods. It achieves notably higher accuracy than other parallel FL and SFL approaches, although slightly lower than sequential SL .

These characteristics and advantages highlight the innovative nature of ASFV in addressing the challenges of vehicular edge computing systems, offering improved efficiency, performance, and adaptability compared to traditional and existing methods in the field.


Do any related researches exist? Who are the noteworthy researchers on this topic in this field?What is the key to the solution mentioned in the paper?

Several related research papers exist in the field of federated learning and vehicular edge computing. Noteworthy researchers in this area include G. Poirot, P. Vepakomma, K. Chang, J. Kalpathy-Cramer, R. Gupta, R. Raskar, W. Wu, M. Li, K. Qu, C. Zhou, X. Shen, W. Zhuang, X. Li, W. Shi, R. Chen, L. Li, K. Xue, C. Zhang, M. Pan, Y. Fang, L. Zhu, F. R. Yu, T. Tang, Z. Zhou, E. Li, L. Zeng, K. Luo, J. Zhang, S. Song, K. B. Letaief, S. Luo, X. Chen, Q. Wu, Z. Zhou, S. Yu, Y. Fu, H. Guo, X. Yang, Y. Ding, V. Chandra, Y. Lin, and many others .

The key to the solution mentioned in the paper involves various strategies such as weight quantization, gradient quantization, joint optimization of local accuracy, transmit power, data rate, and devices' computing capacities, as well as communication-efficient federated learning frameworks with partial model aggregation algorithms and compression strategies. Additionally, split learning (SL) and split federated learning (SFL) have been proposed to address challenges in communication efficiency and training latency, with approaches like Cluster-based Parallel SL (CPSL) and Hybrid Split and Federated Learning (HSFL) being developed to optimize model training mechanisms .


How were the experiments in the paper designed?

The experiments in the paper were designed as follows:

  • The paper proposed a novel low-latency and low-energy Adaptive Split Federated Vehicle (ASFV) training by combining vehicle selection and resource allocation, with a theoretical analysis of training delay and energy consumption .
  • A vehicle selection algorithm based on vehicle speed and EC communication range was formulated, along with a time delay minimization multi-objective function considering vehicle heterogeneity, channel instability, and model splitting strategy .
  • The formulated multi-objective problem was decomposed into three subproblems: online adaptive cut layer selection, transmission power assignment, and wireless resource allocation, which were solved iteratively using BCD, KKT, SCA, and Lagrange multiplier methods .
  • The performance of the proposed solution was evaluated through extensive simulations using various open datasets, demonstrating superior time-energy efficiency and learning performance for ASFV over heterogeneous devices compared to existing schemes .

What is the dataset used for quantitative evaluation? Is the code open source?

The dataset used for quantitative evaluation in the research is a training set with 50,000 samples for model training and a test set with 10,000 samples for performance evaluation . The code for the research is not explicitly mentioned to be open source in the provided context.


Do the experiments and results in the paper provide good support for the scientific hypotheses that need to be verified? Please analyze.

The experiments and results presented in the paper provide substantial support for the scientific hypotheses that needed verification. The paper extensively explores the challenges and advancements in Federated Learning (FL) within Vehicular Edge Computing (VEC) systems, addressing critical issues such as high heterogeneity among vehicles/clients involved in training and the protection of user privacy . The experiments conducted in the paper demonstrate the application of Split Learning (SL) and Federated Learning (FL) in VEC systems, showcasing the potential benefits and challenges associated with these approaches . Additionally, the paper discusses the impact of mobility, communication delays, energy consumption, and resource allocation on the performance of FL and SL algorithms in VEC systems, providing valuable insights into the practical implications of these technologies .

Moreover, the paper introduces innovative approaches such as Split Learning-based Intrusion Detection Systems (IDS) and SplitFed learning with mobility methods to enhance model training efficiency and security in VEC systems . These experiments contribute to the validation of the scientific hypotheses by demonstrating the effectiveness of SL and FL in addressing security concerns and optimizing training times in VEC environments . The results presented in the paper offer a comprehensive analysis of the performance metrics, convergence speeds, and energy consumption associated with FL and SL algorithms in VEC systems, supporting the scientific hypotheses and providing valuable insights for future research in this domain .


What are the contributions of this paper?

The paper makes several contributions in the field of vehicular edge computing and federated learning:

  • Proposing a Split Learning-based IDS (SplitLearn) for ITS infrastructures to address security concerns, outperforming Federated Learning (FedLearn) and Transfer Learning (TransLearn) .
  • Introducing SplitFed learning with a mobility method to minimize model training time and a migration method for the ML model when vehicles move between different VECs .
  • Addressing challenges in Split Learning (SL)-assisted and Split Federated Learning (SFL)-assisted VEC, highlighting the need for detailed analysis on model migration, long SL serial delay, and device allocation for FL and SL .

What work can be continued in depth?

Further research can be conducted to delve deeper into the exploration of Split Learning (SL) as a collaborative learning framework, especially within the context of vehicular networks. While SL has shown promising results in addressing security concerns and outperforming other learning solutions like Federated Learning (FL) and Transfer Learning (TransLearn) , there is still room for in-depth investigation into its application and optimization in vehicular edge computing systems. Specifically, research can focus on enhancing the efficiency and effectiveness of Split Learning in addressing the unique challenges and requirements of vehicular networks, such as optimizing model splitting for supporting a large number of vehicles, considering vehicle heterogeneity, network dynamics, and vehicle mobility . By further exploring adaptive and parallel Split Federated Learning solutions, researchers can aim to optimize performance over wireless networks and improve the overall learning process in vehicular edge computing systems .

Tables
1
Scan the QR code to ask more questions about the paper
© 2025 Powerdrill. All rights reserved.