This diagram illustrates a typical model inference workflow, often seen in collaborative settings like federated learning. Let's break down the key components:
Key Roles
- **Data Owner/Output Owner:** Holds the raw data and may receive the final model for use in a inference pipeline.
- **Model Owner:** Defines and provides the base ML model to a customer domain. It can be used only for finetuning and the output model is used for inference purposes.
Workflow Steps
- **Model Sharing:**
- **ML Base Model:** The initial ML model developed by the Model owner that will be shared to be executed in a customer cloud account. Model owner does not want to lose ownership of the model while in use
- **Data Preparation:**
- **Raw Datasets:** The initial data sources used for finetuning.
- **Extract/Merge/Transform:** Preprocessing the data to extract features, merge sources, and prepare it for training.
- **Transformed Data:** The processed data is ready for tuning the model.
- **Model Tuning:**
- **Data Loader:** Loads and manages the data for training. inside a Confidential Environment
- **Base Model:** The initial model structure and the Shared ML model secure loaded into the Confidential Environment. Encryption in use prevents outside access to the model.
- **Training:** The core process where the model learns from the data by adjusting its parameters.
- **Trained Model:** The final model ready for use in inference. This Model is only available for use inside the inference Confidential Environment.
- **Inference:**
- **Input Data:** New data for the model to make predictions on.
- **Output Inference:** The predictions generated by the model.
Confidential Training & Inference
The labels "Confidential Training" and "Confidential Inference" indicate that the training and inference processes are designed to protect sensitive data, potentially through techniques like Azure Confidential Container Instances.
This workflow highlights the collaborative nature of machine learning and emphasizes the importance of securely sharing models for the purpose of model training and inference. It's particularly relevant in scenarios where data is distributed across organizations, and they want to train a model without sharing raw data directly.