1.5 Roles and Terminology

Taiwan AI Labs Federated Learning Platform (FLP) provides the following main options for AI model implementation:

• FL: Federated Learning
• FV: Federated Validation
• DG: Data Governance

Roles:

• Fed-Admin: Fed Machine Administrator. (地端設備管理者)
  Each on-premises fed machine will have one Taiwan AI Labs authorized Fed-Admin account, and by using this Fed-Admin account, institute's administrator could create new FL/FV/DG projects, create and authorize new PI accounts, and assign the project to the PI.
  
  Fed-Admin will use Aggregator Dashboard to create a new FL/FV project and assign PI.
  Fed-Admin will use Fed Machine to create a new Data Governance (DG) project and assign PI.
  
  
• PI: Principle Investigator. (盟主)
  The lead PI in charge of the whole Federated Learning (FL) project. Each project will have one main PI to take care of the main task, including inviting Co-PI to join the FL/FV project, create FL/FV plans under the project, and upload the pretrained AI model and weights.
  
  For Federated Learning project, PI will use it's own data to generate the AI initial model and initial weights.
  
  PI will use use Aggregator Dashboard to manage the FL/FV projects, including inviting Co-PI, upload AI model and weights, and creating FL/FV plans under FL/FV project.
  
  PI will access Fed Machine to join a FL/FV project for the local edge site.
  PI will use use Fed Dashboard to upload datasets.
  PI will use Edge Machine to manage Data Governance (DG) projects.
  
  
• Co-PI: Co-Principal Investigator. (盟友)
  Co-PI is the PI in other institute, joining the same FL project. Co-PI could join the FL project by the invitation information received from PI. Co-PI will upload it's own datasets to assist the FL project training, or to verify the AI model/weights of FV project.
  
  Co-PI will access Fed Machine to join a FL/FV project, and Fed Dashboard to upload the datasets.
  
  
• Sub-I: Sub-Investigator or Co-Investigator. (助理、助理主持人)
  Sub-I to assist PI in the main institute or Co-PI in the alliance. Sub-I has important roles in Data Governance toolsets of Taiwan AI Labs Federated Learning Platform.

FLP (Federated Learning Platform) Terminology:

• Aggregator Dashboard:
  For each Federated Learning (FL) project, the model weights need to be aggregated in one place, and we name that place as Aggregator, and the managing place is called Aggregator Dashboard. Aggregator Dashboard contains all the functions for PI to manage a FL/FV project. Fed-Admin also creates a new FL project (as well as assigns PI) at Aggregator Dashboard.
  
  Taiwan AI Labs offers the option for users to choose the location of the Aggregator Dashboard. For more information, please contact Taiwan AI Labs directly.
  
  
• Fed Dashboard:
  The dashboard provided at the Fed Machine for PI/Co-PI to operate the FL/FV tasks locally. PI/Co-PI will select the joined FL/FV project, enter the project PIN code, and then login the FL/FV project Fed Dashboard.
  
  
• Fed Machine:
  The full-functions AI machine provided by Taiwan AI Labs. Pre-installed the services provided by Taiwan AI Labs.
  
  Each Fed Machine will also come with a Fed-Admin account, so the institutes could mange the services provided all by itself.
  
  
• Fed Portal:
  Taiwan AI Labs Fed Machine provides a portal web page for users to access all services provided by Taiwan AI Labs.
  Each Fed Machine is associated and set up with institute's internal domain name (FQDN - Fully Quality Domain Name), and you could access this FQDN by using Google Chrome browser directly. The portal page contains all the services' shortcuts offered by Taiwan AI Labs, and this default accessing web page is called Fed Portal.

AI (Artificial Intelligence) Terminology:

• AI Model:
  An mathematical representation of algorithm to perform some specific task.
  
  Here at Taiwan AI Labs Federated Learning Platform, we use Docker technology to communicate with AI model algorithm.
  
  
• AI Model Weights:
  The parameters of the model learned during the training process.
  
  The AI model weights are updated and improved during training at each epoch to minimize the error between the model's predictions and the actual outcomes.
  
  
• Initial AI Model:
  The selected AI model by PI using the datasets at PI's hand. Also named as "Initial Model".
  
  
• Initial Model Weights:
  The AI model weights of the initial AI model generated and optimized by PI's local datasets. Also named as "Initial AI Model Weights" or "Initial Weights".

Machine Learning (ML) and Deep Learning (DL) Terminology:

• Run:
  A run refers to the entire process of training a model.
  A run usually involve multiple rounds, which might contain one or several epochs, each consisting of multiple iterations.
  
  
• Round:
  An round is one time Fed Dashboard (Fed Machine) communicates with Aggregator Dashboard.
  An round starts when Aggregator Dashboard sends training request to Fed Dashboard with last-round aggregated results. Fed Dashboard transfers the re-trained model weights back to the Aggregator Dashboard after one or several epochs, depending on the design. Then one round is completed.
  An round might contain one or multiple epoch .
  
  
• Epoch:
  An epoch refers to a single pass over the entire training dataset during the training process.
  An epoch is completed when the model has seen every example in the training dataset once.
  An epoch consists of multiple iterations.
  
  
• Iteration:
  An iteration refers to a single update of the model's parameters using a batch of training data.
  A batch is a subset of the entire training dataset that is used to update the model's parameters.
  (iteration numbers * batch size = training set size)