Patent attributes
A method includes generating a quadratic penalty based on a first version of a parameter, a first version of the anchor point and a set of Hessian values, training a task using the quadratic penalty, determining a second version of the parameter based on the trained task, determining a second version of the anchor point for the task based a trained task, generating Hessian values for the task based on the trained task, calculating a loss value for a portion of the plurality of tasks based on the second version of the parameter, calculating a loss value for each of the remainder of the plurality of tasks, determining if the loss value satisfies a condition, and in response to determining the loss value satisfies the condition, replace the first version of the parameter with the second version of the parameter in a parameter datastore and in the version control datastore.