ISLAB/CAISR - User contributions [en]

Model Heterogeneity in Federated Learning

2020-10-08T10:16:12Z

Amira:

{{StudentProjectTemplate
|Summary=Group users within a federated learning environment into different learning overlays according to their behavioural similarities
|Keywords=Federated Learning, Clustering
|References=Advances and Open Problems in Federated Learning:
https://hal.inria.fr/hal-02406503/document

FedML: A Research Library and Benchmark for Federated Machine Learning: https://arxiv.org/pdf/2007.13518.pdf

|Supervisor=Amira Soliman, Sławomir Nowaczyk,
|Level=Master
|Status=Open
}}
Federated Learning (FL) has been introduced as an alternative distributed and privacy-friendly learning approach. FL allows users to train models locally on their devices using their sensitive data, and communicate intermediate model updates to a central server without the need to centrally store the data. The principal advantage of FL is the decoupling of global model training from the need for direct access to the raw data. Accordingly, FL offers a solution to learn from private personal data such as biometrics, text input, and location coordinates where models can be trained for many services in a privacy-preserving manner.

Generating a single global model that accumulates all user behaviors might not produce the best model for particular categories of users. Specifically, the global averaging model enforces a bias towards the behavioral patterns provided by the majority, while suppressing the patterns of less significant users. Thus, it is interesting to provide overlay-based FL techniques that can group users in different learning overlays according to their behavioral similarities. The objective of this thesis is to introduce a mechanism for grouping users with similar behaviors and develop a hierarchical aggregation mechanism to provide more than one model, automatically identifying the best group for a given node.

Data Heterogeneity in Federated Learning

2020-10-08T10:16:00Z

Amira:

{{StudentProjectTemplate
|Summary=Addressing the challenges of data imbalance in Federated Learning
|Keywords=non-IID data, Federated Learning
|References=Advances and Open Problems in Federated Learning:
https://hal.inria.fr/hal-02406503/document

FedML: A Research Library and Benchmark for Federated Machine Learning: https://arxiv.org/pdf/2007.13518.pdf

|Supervisor=Amira Soliman, Slawomir Nowaczyk
|Level=Master
|Status=Open
}}
Federated Learning (FL) has been introduced as an alternative distributed and privacy-friendly learning approach. FL allows users to train models locally on their devices using their sensitive data, and communicate intermediate model updates to a central server without the need to centrally store the data. The principal advantage of FL is the decoupling of global model training from the need for direct access to the raw data. Accordingly, FL offers a solution to learn from private personal data such as biometrics, text input, and location coordinates where models can be trained for many services in a privacy-preserving manner.

A common assumption in FL is that each node has an unbiased sample of the complete data. In reality, though, the models created by different users can often be quite different, as the data on each device can originate from different phenomena. For example, two randomly picked users are likely to compute very different updates to a typing prediction model. This leads to a situation that is challenging from the statistical standpoint, and most existing methods make strong assumptions for how skewed the data distributions are. However, in open and decentralized environments, imbalanced data as well as missing classes are common, and it is imperative that FL methods can deal with them. A lot of work has been done for solving class imbalance and missing classes in a centralized setting, but it is more challenging to provide practical and privacy-preserving learning methods for imbalanced data in FL environments. The objective of this thesis is to create an FL algorithm that is capable of handling the data imbalance among participating devices and propose a solution to enhance model training under this imbalance.

Data Heterogeneity in Federated Learning

2020-10-08T10:14:31Z

Amira:

{{StudentProjectTemplate
|Summary=Addressing the challenges of data imbalance in Federated Learning
|Keywords=non-IID data, Federated Learning
|Supervisor=Amira Soliman, Slawomir Nowaczyk
|Level=Master
|Status=Open
}}
Federated Learning (FL) has been introduced as an alternative distributed and privacy-friendly learning approach. FL allows users to train models locally on their devices using their sensitive data, and communicate intermediate model updates to a central server without the need to centrally store the data. The principal advantage of FL is the decoupling of global model training from the need for direct access to the raw data. Accordingly, FL offers a solution to learn from private personal data such as biometrics, text input, and location coordinates where models can be trained for many services in a privacy-preserving manner.

A common assumption in FL is that each node has an unbiased sample of the complete data. In reality, though, the models created by different users can often be quite different, as the data on each device can originate from different phenomena. For example, two randomly picked users are likely to compute very different updates to a typing prediction model. This leads to a situation that is challenging from the statistical standpoint, and most existing methods make strong assumptions for how skewed the data distributions are. However, in open and decentralized environments, imbalanced data as well as missing classes are common, and it is imperative that FL methods can deal with them. A lot of work has been done for solving class imbalance and missing classes in a centralized setting, but it is more challenging to provide practical and privacy-preserving learning methods for imbalanced data in FL environments. The objective of this thesis is to create an FL algorithm that is capable of handling the data imbalance among participating devices and propose a solution to enhance model training under this imbalance.

Model Heterogeneity in Federated Learning

2020-10-08T10:14:10Z

Amira:

{{StudentProjectTemplate
|Summary=Group users within a federated learning environment into different learning overlays according to their behavioural similarities
|Keywords=Federated Learning, Clustering
|Supervisor=Amira Soliman, Sławomir Nowaczyk,
|Level=Master
|Status=Open
}}
Federated Learning (FL) has been introduced as an alternative distributed and privacy-friendly learning approach. FL allows users to train models locally on their devices using their sensitive data, and communicate intermediate model updates to a central server without the need to centrally store the data. The principal advantage of FL is the decoupling of global model training from the need for direct access to the raw data. Accordingly, FL offers a solution to learn from private personal data such as biometrics, text input, and location coordinates where models can be trained for many services in a privacy-preserving manner.

Generating a single global model that accumulates all user behaviors might not produce the best model for particular categories of users. Specifically, the global averaging model enforces a bias towards the behavioral patterns provided by the majority, while suppressing the patterns of less significant users. Thus, it is interesting to provide overlay-based FL techniques that can group users in different learning overlays according to their behavioral similarities. The objective of this thesis is to introduce a mechanism for grouping users with similar behaviors and develop a hierarchical aggregation mechanism to provide more than one model, automatically identifying the best group for a given node.

Model Heterogeneity in Federated Learning

2020-10-08T09:46:25Z

Amira:

{{StudentProjectTemplate
|Summary=Group users within a federated learning environment into different learning overlays according to their behavioural similarities
|Keywords=Federated Learning, Clustering
|Supervisor=Amira Soliman, Sławomir Nowaczyk,
}}
Federated Learning (FL) has been introduced as an alternative distributed and privacy-friendly learning approach. FL allows users to train models locally on their devices using their sensitive data, and communicate intermediate model updates to a central server without the need to centrally store the data. The principal advantage of FL is the decoupling of global model training from the need for direct access to the raw data. Accordingly, FL offers a solution to learn from private personal data such as biometrics, text input, and location coordinates where models can be trained for many services in a privacy-preserving manner.

Generating a single global model that accumulates all user behaviors might not produce the best model for particular categories of users. Specifically, the global averaging model enforces a bias towards the behavioral patterns provided by the majority, while suppressing the patterns of less significant users. Thus, it is interesting to provide overlay-based FL techniques that can group users in different learning overlays according to their behavioral similarities. The objective of this thesis is to introduce a mechanism for grouping users with similar behaviors and develop a hierarchical aggregation mechanism to provide more than one model, automatically identifying the best group for a given node.

Model Heterogeneity in Federated Learning

2020-10-08T09:45:34Z

Amira: Created page with "{{StudentProjectTemplate |Summary=Group users within a federated learning environment into different learning overlays according to their behavioural similarities |Keywords=Fe..."

{{StudentProjectTemplate
|Summary=Group users within a federated learning environment into different learning overlays according to their behavioural similarities
|Keywords=Federated Learning, Clustering
|Supervisor=Amira Soliman, Sławomir Nowaczyk,
}}
Federated Learning (FL) has been introduced as an alternative distributed and privacy-friendly learning approach. FL allows users to train models locally on their devices using their sensitive data, and communicate intermediate model updates to a central server without the need to centrally store the data. The principal advantage of FL is the decoupling of global model training from the need for direct access to the raw data. Accordingly, FL offers a solution to learn from private personal data such as biometrics, text input, and location coordinates where models can be trained for many services in a privacy-preserving manner.

Generating a single global model that accumulates all user behaviors might not produce the best model for particular categories of users. Specifically, the global averaging model enforces a bias towards the behavioral patterns provided by the majority, while suppressing the patterns of less significant users. Thus, it is interesting to provide overlay-based FL techniques that can group users in different learning overlays according to their behavioral similarities. The objective of this thesis is to introduce a mechanism for grouping users with similar behaviors and develop a hierarchical aggregation mechanism to provide more than one model, automatically identifying the best group for a given node.

Data Heterogeneity in Federated Learning

2020-10-08T09:41:02Z

Amira: Created page with "{{StudentProjectTemplate |Summary=Addressing the challenges of data imbalance in Federated Learning |Keywords=non-IID data, Federated Learning |Supervisor=Amira Soliman, Slawo..."

{{StudentProjectTemplate
|Summary=Addressing the challenges of data imbalance in Federated Learning
|Keywords=non-IID data, Federated Learning
|Supervisor=Amira Soliman, Slawomir Nowaczyk
}}
Federated Learning (FL) has been introduced as an alternative distributed and privacy-friendly learning approach. FL allows users to train models locally on their devices using their sensitive data, and communicate intermediate model updates to a central server without the need to centrally store the data. The principal advantage of FL is the decoupling of global model training from the need for direct access to the raw data. Accordingly, FL offers a solution to learn from private personal data such as biometrics, text input, and location coordinates where models can be trained for many services in a privacy-preserving manner.

A common assumption in FL is that each node has an unbiased sample of the complete data. In reality, though, the models created by different users can often be quite different, as the data on each device can originate from different phenomena. For example, two randomly picked users are likely to compute very different updates to a typing prediction model. This leads to a situation that is challenging from the statistical standpoint, and most existing methods make strong assumptions for how skewed the data distributions are. However, in open and decentralized environments, imbalanced data as well as missing classes are common, and it is imperative that FL methods can deal with them. A lot of work has been done for solving class imbalance and missing classes in a centralized setting, but it is more challenging to provide practical and privacy-preserving learning methods for imbalanced data in FL environments. The objective of this thesis is to create an FL algorithm that is capable of handling the data imbalance among participating devices and propose a solution to enhance model training under this imbalance.

Feature-wise normalization for 3D medical images

2020-09-29T13:40:00Z

Amira:

{{StudentProjectTemplate
|Summary=Normalization of 3D medical imaging either as a data pre-processing or as feature-wise batch normalization during CNN model training
|Keywords=CNN, 3D models
|Supervisor=Amira Soliman, Stefan Byttner, Kobra Etminani
|Level=Master
|Status=Open
}}
Normalization is a required preprocessing step, especially for deep learning and convolutional neural networks, such that the network becomes unbiased towards the different features. However, in medical images, the whole intensity normalization may lead to reduced sensitivity for relatively important features. The objective of this master thesis is to study the state-of-the-art normalization techniques used in 2D images, investigate the applicability of such techniques in 3D medical images, and apply them either as a preprocessing step or as feature-wise batch normalization during the model training.

Prioritize informative structures in 3D brain images

2020-09-29T13:39:34Z

Amira:

{{StudentProjectTemplate
|Summary=Identify informative regions in 3D brain images to improve classification accuracy of dementia disorders
|Keywords=CNN, 3D models
|Supervisor=Amira Soliman, Kobra Etminiani, Stefan Byttner,
|Level=Master
|Status=Open
}}
3D PET scans show 3D images of the cell activity in the tissues of the human brain. Having these scans, doctors can use the computer-aided diagnosis of dementia disorders like Alzheimer’s and Parkinson's. 3D PET scans are considered as high dimensional data, though not all of the layers are used during the analysis of such data, especially within classification tasks. Furthermore, the automatic classification using 3D brain images can be applied to the whole brain or using specific regions of interest (ROIs) that can be considered as structure biomarkers and relate them to particular dementia disorders. The objective of this thesis is to investigate extracting informative regions across the different 3D layers of PET scans and assess the contribution of such regions to the classification accuracy. Identifying such regions can be performed with the help of extra domain knowledge or brain parcellation methods.

Feature-wise normalization for 3D medical images

2020-09-29T13:38:10Z

Amira:

{{StudentProjectTemplate
|Summary=Normalization of 3D medical imaging either as a data pre-processing or as feature-wise batch normalization during CNN model training
|Keywords=CNN, 3D models
|TimeFrame=2020 Fall - 2021 Summer
|Prerequisites=Excellent Programming Skills
Excellent knowledge in Machine Learning and Neural Networks
|Supervisor=Amira Soliman, Stefan Byttner, Kobra Etminani
|Level=Master
|Status=Open
}}
Normalization is a required preprocessing step, especially for deep learning and convolutional neural networks, such that the network becomes unbiased towards the different features. However, in medical images, the whole intensity normalization may lead to reduced sensitivity for relatively important features. The objective of this master thesis is to study the state-of-the-art normalization techniques used in 2D images, investigate the applicability of such techniques in 3D medical images, and apply them either as a preprocessing step or as feature-wise batch normalization during the model training.

Prioritize informative structures in 3D brain images

2020-09-29T13:36:57Z

Amira:

Name of the new project

2020-09-29T13:35:30Z

Amira: Amira moved page Name of the new project to Feature-wise normalization for 3D medical images

#REDIRECT [[Feature-wise normalization for 3D medical images]]

Feature-wise normalization for 3D medical images

2020-09-29T13:35:30Z

Amira: Amira moved page Name of the new project to Feature-wise normalization for 3D medical images

{{StudentProjectTemplate
|Summary=Normalization of 3D medical imaging either as a data reprocessing or as feature-wise batch normalization during CNN model training
|Keywords=CNN, 3D models
|TimeFrame=2020 Fall - 2021 Summer
|Prerequisites=Excellent Programming Skills
Excellent knowledge in Machine Learning and Neural Networks
|Supervisor=Amira Soliman, Stefan Byttner, Kobra Etminani
|Level=Master
|Status=Open
}}
Normalization is a required preprocessing step, especially for deep learning and convolutional neural networks, such that the network becomes unbiased towards the different features. However, in medical images, the whole intensity normalization may lead to reduced sensitivity for relatively important features. The objective of this master thesis is to study the state-of-the-art normalization techniques used in 2D images, investigate the applicability of such techniques in 3D medical images, and apply them either as a preprocessing step or as feature-wise batch normalization during the model training.

Prioritize informative structures in 3D brain images

2020-09-29T13:32:40Z

Amira: Created page with "{{StudentProjectTemplate |Summary=Identify informative regions in 3D brain images |Keywords=CNN, 3D models |Supervisor=Amira Soliman, Kobra Etminiani |Level=Master |Status=Ope..."

{{StudentProjectTemplate
|Summary=Identify informative regions in 3D brain images
|Keywords=CNN, 3D models
|Supervisor=Amira Soliman, Kobra Etminiani
|Level=Master
|Status=Open
}}
3D PET scans show 3D images of the cell activity in the tissues of the human brain. Having these scans, doctors can use the computer-aided diagnosis of dementia disorders like Alzheimer’s and Parkinson's. 3D PET scans are considered as high dimensional data, though not all of the layers are used during the analysis of such data, especially within classification tasks. Furthermore, the automatic classification using 3D brain images can be applied to the whole brain or using specific regions of interest (ROIs) that can be considered as structure biomarkers and relate them to particular dementia disorders. The objective of this thesis is to investigate extracting informative regions across the different 3D layers of PET scans and assess the contribution of such regions to the classification accuracy. Identifying such regions can be performed with the help of extra domain knowledge or brain parcellation methods.

Feature-wise normalization for 3D medical images

2020-09-29T13:29:49Z

Amira: