Incremental Learning Object Detector Prevents Catastrophic Forgetting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Neural network object detectors face catastrophic forgetting when retrained for new object classes, leading to a loss of ability to detect previously learned classes, and require significant computational resources or access to large datasets for updates.
Innovation Solution
A method for incremental learning that uses a total loss function to balance the detection of new and old classes without requiring access to original training data, allowing for iterative updates of a two-stage neural network object detector to maintain detection accuracy for both old and new classes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If incremental learning is used to retrain an object detector for new object classes, then the detector can detect new classes, but it loses the ability to detect previously learned classes (catastrophic forgetting)
Solution Approach 1:
The method performs preliminary actions by computing and storing loss values for previously learned classes before training on new classes. These loss values are calculated using the current model weights and stored for later use during incremental training, ensuring that the model's performance on old classes is preserved without requiring access to original training data.
Solution Approach 2:
The method implements feedback by using the stored loss values from previously learned classes as a constraint during incremental training. The total loss function combines the loss for new classes with the stored loss values for old classes, providing continuous feedback that prevents catastrophic forgetting while enabling learning of new object classes.
2Reliability
If a new object detector is run in parallel with the existing detector to avoid catastrophic forgetting, then both old and new classes can be detected, but computational resources (memory and execution capacity) are doubled
Solution Approach 1:
The method merges the detection capabilities for old and new classes into a single object detector model. By combining the loss functions and training processes, the model incrementally learns new classes while preserving performance on old classes, avoiding the need to run separate detectors in parallel and thus reducing computational resource requirements.
3Measurement precision
If the object detector is trained with a large dataset including both new and previously learned classes, then detection accuracy for both classes improves, but computational resources are dramatically increased
Solution Approach 1:
The method extracts and stores only the essential information (loss values) from the training process for previously learned classes, rather than retaining the entire training dataset. This extracted loss information is sufficient to prevent catastrophic forgetting during incremental training, significantly reducing the computational resources and data storage requirements while maintaining detection accuracy.
Data Source
AI summary
Methods and systems perform incremental learning object detection in images and/or videos without catastrophic forgetting of previously-learned object classes. A two-stage neural network object detector is trained to locate and identify objects pertaining to an additional object class by iteratively updating the two-stage neural network object detector until an overall detection accuracy criterion is met. The updating is performed so as to balance minimizing a loss of an initial ability to locate and identify objects pertaining to the previously-learned object classes and maximizing an ability to additionally locate and identify the objects pertaining to the additional object class. Assessing whether the overall detection accuracy criterion is met compares outputs of an initial version of the two-stage neural network object detector with a current region proposal output by a current version of the two-stage neural network object detector to determining a region proposal distillation loss and a previously-learned-object identification distillation loss.