Models
Similar to the Datasets, training an AI model consists of two steps:
- Model configuration: Defines the configuration for training, such as which dataset to use or whether to retrain automatically.
- Models: The actual trained AI model, ready to be deployed.
Create a model configuration
To access and create model configs, navigate to AI > Model.
Step 1: Architecture
Select an AI model architecture. These options vary in complexity and processing power:
- Iris Nano: Fast, efficient architecture for general-purpose anomaly detection on small datasets. Supports up to 20 items.
- Iris Core: Robust architecture for moderately sized datasets with lower training variance. Supports up to 40 items.
- Iris Pro: Advanced architecture for large datasets with superior detection capabilities. Supports up to 80 items.
- Iris Ultra: The most powerful architecture for highly complex systems. Supports up to 128 items.
Step 2: Configuration
- Name: Name of the model config.
- Dataset config: Dataset config to access a dataset.
- Dataset: Dataset of the provided dataset config.
- Dataset duration mode: Choose Unlimited or Limited duration to manage your storage space automatically.
Only available to companies with dedicated AI trainers:
- AI Trainer: In case your company has a dedicated AI trainer, you can specify what AI trainer to use for model training.
Managing model configurations
Click on any model configuration in the list to view its details. The page is organized into four key sections:
General
Review your selected AI model architecture and update your Name or Description for this AI model configuration.
Additionally, Automatic retraining can be enabled to keep your AI models up to date. When enabled, Prisma will automatically start a training a fresh model when a new dataset is generated by your linked Dataset configuration.
Dataset
This section links your AI to its data source:
- Dataset configuration: Select the dataset configuration you want to use for AI model training.
- Dataset: Choose the specific extracted dataset you want ot use for the training.
Storage options
Review or adjust your Storage duration mode settings.
Training an AI model
To manually start the training process:
- Ensure you have selected a valid dataset in the Dataset section.
- Scroll to the Models section.
- Click Train model.
Once training begins, the system will start training your AI model based on your selected dataset. Click on Details on the new AI model to view the training progress in detail.
Model details and training progress
The AI model details page provides a real-time view of the models development.
Model configuration
In the Model configuration section, you can view the AI model architecture and the Dataset that was used to train this AI model.
Training
In the Training section, you can see the current status of the AI model training (e.g., Training in progress, Training successfully completed, Training failed).
When the AI model is currently training, you can view the progress of the training in real-time:
- Training progress: A live percentage indicator shows how much of the data has been processed.
- Training graph: This visualizes the learning curve of the AI.
- Job log: For full transparency, the log records every event during the training process, from data loading to final optimization.
Model testing
Before deploying a model to your live environment, you can perform a Model Test. This allows you to see how the AI model would have reacted to historical data it has never seen before.
Create a model test
To ensure a fair test, you should select a dataset that is different from the one used during training.
- On the model details page, scroll to the Model tests section.
- Click on Create new test.
- Enter a name for the model test and select a dataset.
- Click Create.
Evaluating test results
Click on the created AI model test to view its progress and details.
The system processes the test data exactly as it would in a real-world deployment. The model evaluates the data and generates Anomaly scores.
- Anomaly graph: This graph shows anomaly scores of the test dataset.
- Anomaly markers: Any points where the AI detected unusual behavior are clearly marked on the graph.