Continual Learning Training With Data Essence Memory
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Catastrophic forgetting in AI/ML models leads to a decline in overall classification accuracy due to the inability to maintain performance on both old and new data, with existing methods like retaining important data or encoding old data being ineffective in continual learning scenarios.
Innovation Solution
A training method for a continual learning model that initializes a replay memory and essence memory, trains an encoder and self-attention layer, and converts raw data into data essence, storing it in the essence memory, while allowing the use of publicly available models and fine-tuning parameters during the first task to adapt to the current task.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a model is trained with new data, then the model learns new tasks, but the model forgets previously learned data (catastrophic forgetting)
Solution Approach 1:
The patent applies preliminary action by pre-processing old data into compressed data essences before they are needed for training. The essence extraction is performed in advance, allowing the model to quickly retrieve and utilize essential features of old tasks without accessing the full original datasets, thus preventing catastrophic forgetting while making room for new task learning.
Solution Approach 2:
The patent extracts the essential features and characteristics of old task data, separating them from the complete original datasets. This extraction process creates compressed data essences that contain the critical information needed for maintaining performance on old tasks, while the full datasets can be discarded or archived, resolving the contradiction between retaining old knowledge and learning new tasks.
2Reliability
If old data is retained and included during training with new data, then performance loss is mitigated, but training data requirements increase
Solution Approach 1:
The patent extracts only the essential features from old task data, creating compressed data essences that contain the critical information needed for maintaining performance. This extraction allows the system to retain the necessary knowledge from old tasks without storing or processing the complete original datasets, thus reducing training data requirements while mitigating performance loss.
Solution Approach 2:
The patent segments the training data into two distinct components: compressed data essences representing old tasks and full datasets for new tasks. This segmentation allows the model to efficiently utilize essential information from old tasks through the compact essences while processing new task data in full, optimizing both performance and data management.
3Measurement precision
If encoding model is pre-trained on target data domain, then encoding effectiveness improves, but preparation time increases
Solution Approach 1:
The patent implements self-service by enabling the encoding model to automatically adapt to new data domains without requiring manual pre-training or configuration. The model performs essence extraction directly on incoming data streams, self-adjusting to different domains as they appear, thus improving encoding effectiveness while eliminating time-consuming pre-preparation steps.
Data Source
AI summary
A training method for continual learning model and a non-transitory computer-readable medium are proposed. The method includes: training the encoder and self-attention layer in the essence generation procedure according to the raw data of a task when the current training process is the first task in continual learning; otherwise, freezing the parameters of the encoder and self-attention layer, performing the essence generation procedure to convert the raw data into a data essence, and adding the data essence into the essence memory. The training process is repeated until the continual learning model converges. The training process includes: obtaining a training batch from the raw data, updating the replay memory according to the training batch, training the continual learning model according to the replay memory and the essence memory, and updating the data essence in the essence memory when the current training process is the first task.