In your finished task no. 2, how did you validate the performance (i.e., preservation of diversity among models) of your ensemble distillation algorithm? What training and testing data did you use?
Will you attempt to apply your model for a specific application, e.g., medical diagnosis, etc.?
How did you validate the performance?
We have designed our version of the FedEED algorithm and have implemented it in code. We are currently running a simulation based on the modified FedEED algorithm. The current round is 61/100. Therefore, we don’t have the results of our simulation yet.
The performance is planned to be validated based on two key criteria, accuracy on highly non-IID data and the number of communication rounds. However, we have not yet devised an idea to quantify ‘diversity.’ Thanks to your question, we will now devise a method to quantify the degree of ‘diversity’ of E-FedEED algorithms.
What training and testing data did you use?
The training data currently in use is CIFAR10 for client-side training and CIFAR100 for serverside distillation. The validation dataset (test dataset) that will be used for the current simulation is CIFAR10. Whilst the test dataset plan to remain constant, we plan to try to experiment with extensive training datasets, including MNIST, EMNIST, and Imagenet.
Will you attempt to apply your model for a specific application, e.g., medical diagnosis, etc.?
No. We do not have a plan yet. However, we wanted to test our algorithm and the outcome model on sensor data if time permits. We would also love to test our algorithm with medical images, as federated learning has massive potential with applications related to medical images.
Thank you.
Best,
Hangyul Son
SON, Hangyul
January 18, 2023 8:03 am
At the moment of writing this comment, our group has made further progress than stated on the poster. This includes,
Designed experiment methodologySet up an experiment simulation environment (FedDF, FedEED)Give an attempt to modify the FedEED algorithm as the first step to designing an E-FedEED algorithm.
Last edited 1 year ago by SON, Hangyul
SON, Hangyul
January 18, 2023 8:03 am
One clarification our group wants to make is that we plan to build an improved algorithm over the FedEED algorithm, which already exists. The improved version of the FedEED algorithm would be named as E-FedEED algorithm, short for extended FedEED. The clarification has been made on the progress report but not on the poster. We apologize for causing any misunderstanding.
Dear students,
I have a couple of questions:
Dear Professor Quadeer,
Thank you for your interest in our project!
How did you validate the performance?
We have designed our version of the FedEED algorithm and have implemented it in code. We are currently running a simulation based on the modified FedEED algorithm. The current round is 61/100. Therefore, we don’t have the results of our simulation yet.
The performance is planned to be validated based on two key criteria, accuracy on highly non-IID data and the number of communication rounds. However, we have not yet devised an idea to quantify ‘diversity.’ Thanks to your question, we will now devise a method to quantify the degree of ‘diversity’ of E-FedEED algorithms.
What training and testing data did you use?
The training data currently in use is CIFAR10 for client-side training and CIFAR100 for serverside distillation. The validation dataset (test dataset) that will be used for the current simulation is CIFAR10. Whilst the test dataset plan to remain constant, we plan to try to experiment with extensive training datasets, including MNIST, EMNIST, and Imagenet.
Will you attempt to apply your model for a specific application, e.g., medical diagnosis, etc.?
No. We do not have a plan yet. However, we wanted to test our algorithm and the outcome model on sensor data if time permits. We would also love to test our algorithm with medical images, as federated learning has massive potential with applications related to medical images.
Thank you.
Best,
Hangyul Son
At the moment of writing this comment, our group has made further progress than stated on the poster. This includes,
Designed experiment methodologySet up an experiment simulation environment (FedDF, FedEED)Give an attempt to modify the FedEED algorithm as the first step to designing an E-FedEED algorithm.
One clarification our group wants to make is that we plan to build an improved algorithm over the FedEED algorithm, which already exists. The improved version of the FedEED algorithm would be named as E-FedEED algorithm, short for extended FedEED. The clarification has been made on the progress report but not on the poster. We apologize for causing any misunderstanding.