1.3. Collaborative Learning Scenario¶
In this Scenario, Model Manager Receiver and Model Manager Sender nodes working on the same problem share their locally obtained models with each other, without having to share the private datasets with which they were trained on. This intends to lead towards a more complex and accurate model. It leverages the RPC over DDS communication protocol/paradigm in order to exchange all required information (model requests/replies) in an efficient way.
The Model Manager Sender Nodes publish Model Statistics Data Type ‘s while the Model Manager Receiver Nodes listen to them. When a Model Manager Receiver Node is interested in a model based on its Model Statistics Data Type, it sends a Model Request Data Type request to the Model Manager Sender Node that sent the Model Statistics Data Type. The Model Manager Sender Node will respond with a Model Reply Data Type.
1.3.1. Model Request Data Type¶
The Model Request Data Type represents a problem request.
Internally, requests sent from a Model Manager Receiver Node to a Model Manager Sender Node are treated as byte arrays of arbitrary size.
So far, the interaction with this class could be done from a void*, a byte array or a string.
Note
A more specific Data Type will be implemented in future releases for efficiency improvements.
1.3.2. Model Reply Data Type¶
The Model Reply Data Type represents a problem reply with the requested model.
The replies sent from a Model Manager Sender Node to a Model Manager Receiver Node are treated as a bytes array of arbitrary size.
So far, the interaction with this class could be done from a void*, a byte array or a string.
Note
A more specific Data Type will be implemented in future releases for efficiency improvements.
1.3.3. Model Statistics Data Type¶
The Statistics Data Type represents the statistics of models, such as their number of parameters or the datasets they were trained on.
The messages sent from a Model Manager Sender Node to a Model Manager Receiver Node are treated as a bytes array of arbitrary size.
So far, the interaction with this class could be done from a void*, a byte array or a string.
Note
A more specific Data Type will be implemented in future releases for efficiency improvements.