Evolution

Track changes over time

The platform includes a powerful feature for tracking changes between training sessions. With the click of a button, users can view a detailed comparison, highlighting the differences in raw JSON, of any two training sessions simultaneously.

This feature provides users with ground truth on specific changes made between runs, facilitating the evolution of models over time. By making it easy to track these changes.

Transfer Learning

Transfer learning leverages the knowledge gained from prior training sessions, which can accelerate the convergence of new models and improve performance, especially when working with complex data or limited resources. This method is particularly useful for fine-tuning models, allowing users to adapt a pre-trained model to new tasks or datasets with minimal additional training.

When launching a training session, users can select a previously trained model to load weights from, enabling the practice of transfer learning.

Every node on the canvas has a component name. When transfer learning is enabled, the platform matches up neurons by their component name as guide for loading weights into the correct neurons.

Every node on the canvas as a property named is_trainable. By default, this value is true. If the canvas author overrides this and sets it to false, then all of the neurons belonging to that node in the neural network will have their values locked.

With these two simple mechanisms, the platform offers an unprecedented amount of control for a GUI.