Data Essence Replay Training to Reduce Catastrophic Forgetting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Catastrophic forgetting occurs in AI/ML models when they are continually trained on new data, leading to decreased performance on both new and old data, with traditional mitigation approaches being insufficiently effective.
Innovation Solution
A method involving data essence generation, storage in an essence memory, and repeated training procedures is employed, where data essence is generated from raw data, stored, and used to update a model's replay memory, simulating human memory retention by preserving important data details.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If the model is continually trained on new data, then the model's ability to learn new patterns is improved, but the model's performance on old data deteriorates due to catastrophic forgetting
Solution Approach 1:
The patent creates a synthetic copy of important data patterns through the essence generation process. The essence generator produces artificial data representations that capture the salient features of training data, which are then used to train a second model. This copy mechanism allows the system to preserve knowledge of old data patterns while learning new data, resolving the contradiction between adapting to new patterns and maintaining performance on old data.
Solution Approach 2:
The patent introduces an intermediary essence generator that bridges the gap between original training data and the model training process. This intermediary component extracts and transforms important data patterns into an essence representation, which then serves as the training material for the second model. This intermediary layer prevents direct conflict between old and new data, allowing both to coexist without catastrophic forgetting.
2Reliability
If traditional approaches retain a small amount of important data, then performance degradation is reduced, but the effectiveness is limited and does not sufficiently address catastrophic forgetting
Solution Approach 1:
The patent fundamentally changes the parameter of data representation from raw data to essence-generated synthetic data. Instead of retaining and using original training data directly, the system transforms data into an essence representation that captures its essential patterns. This parameter change enables more effective mitigation of catastrophic forgetting by creating a compressed, essential representation that can be efficiently reused during training.
Solution Approach 2:
The patent performs preliminary action by pre-training an essence generator on the training data before the main training process. This preliminary essence generation creates a library of synthetic data representations that can be drawn from during subsequent training. This advance preparation ensures that important patterns are captured before they are needed, making the mitigation of catastrophic forgetting more effective.
Data Source
AI summary
A method of training a model using data essence is performed by a computing device and includes: performing an essence generating procedure according to raw datum to generate a data essence, adding the data essence to an essence memory, and repeatedly performing a training procedure before the model converges. The training procedure includes: obtaining a training batch, updating a replay memory according to the training batch, wherein the replay memory before updating includes a plurality of data from an old training batch, and training the model according to the replay memory and the essence memory.