Neural Network Task Expansion While Preserving Learned Features

The patent applies preliminary action by pre-training the model on source tasks to acquire useful feature representations before adding new tasks. The pre-trained parameters serve as a foundation that can be incrementally updated when new tasks are introduced, avoiding complete random initialization and preserving previously learned features.

The patent employs parameter changes by dynamically adjusting the initialization strategy for new task parameters. Instead of always using random initialization, the system selectively initializes parameters based on the relationship between source and target tasks, allowing for smoother transitions and better preservation of learned features while maintaining task flexibility.

The patent applies local quality by differentiating the training approach for different parts of the model and different tasks. Source tasks use a training approach optimized for their specific requirements, while target tasks use an approach that leverages transferred features. This allows each task to receive appropriate training attention while sharing common feature representations.

The patent segments the task learning process into source task training and target task transfer learning. By separating the training phases and applying different optimization strategies to each segment, the system can maintain high accuracy for source tasks while effectively adapting to new target tasks without compromising overall performance.