Backbone Network Weight Selection Cycle for Training Efficiency

The patent segments the weights in the backbone network into multiple groups based on their importance or characteristics. Instead of adjusting all weights uniformly, the method selectively adjusts only certain weight groups in each training cycle. This segmentation allows the training process to focus on critical weights, reducing the overall computational burden while maintaining training effectiveness.

The patent applies partial action by adjusting only a subset of weights rather than all weights during each training cycle. The method determines which weights require adjustment based on predefined criteria or importance metrics, and selectively updates only those weights. This partial adjustment strategy reduces the number of operations per training cycle, thereby improving training efficiency while still achieving reliable model performance.

The patent implements periodic action by using a weight selection cycle that alternates between different training strategies. Within each cycle, the method periodically selects and adjusts weights based on accumulated gradient information or other metrics. This periodic approach allows the system to balance between frequent updates (for speed) and thorough evaluation (for accuracy), as the selection criteria are applied at regular intervals rather than continuously.

The patent applies preliminary action by pre-calculating or pre-identifying which weights should be adjusted before the actual training cycle begins. The method uses preliminary metrics such as gradient magnitude or weight sensitivity analysis to determine the set of weights to be updated, allowing the main training process to proceed efficiently without real-time decision-making overhead that would slow down training.