Selective Mask Weight Updates for Edge Incremental Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing AI systems in edge devices face challenges in efficiently updating models with new tasks due to high computational and memory requirements, particularly in gradient descent-based incremental learning processes, which are not optimized for edge device resources.
Innovation Solution
A selective weight update (SWU) method that identifies and updates only targeted mask weights in a pretrained model using binary masks and learning rate adjustments, reducing unnecessary computations and memory access.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If gradient descent algorithm-based incremental learning is used to train a model with new tasks, then the model can effectively learn new classes, but the computation and memory access requirements become extremely high making direct use difficult in edge devices
Solution Approach 1:
The patent segments the weight update process by introducing a binary mask that selectively activates only certain weights for updating. Instead of updating all weights in the pretrained model during incremental learning, the binary mask divides weights into updateable and non-updateable portions, reducing the computational scope while maintaining learning effectiveness on new tasks.
Solution Approach 2:
The patent applies local quality by allowing different weights to have different update characteristics through the binary mask. Each weight can be independently marked as updateable or non-updateable based on its importance, enabling localized updates only where necessary rather than uniform updates across the entire model, thus reducing overall computation while preserving critical learned features.
2Adaptability or versatility
If all weights in the pretrained model are updated during incremental learning, then the model can adapt to new tasks, but the memory access and computation requirements become prohibitive for edge devices
Solution Approach 1:
The weight set is segmented into updateable and non-updateable portions using a binary mask. This segmentation reduces the effective model size that needs to be processed during incremental learning, making the computation and memory access manageable for edge devices while still allowing the model to adapt to new tasks through selective weight updates.
Solution Approach 2:
The patent extracts only the essential weights that need to be updated for new task adaptation, separating them from the pretrained weights that should remain fixed. By taking out only the necessary updateable weights and protecting the rest through the binary mask, the system reduces memory access requirements and computational complexity while maintaining adaptability.
3Productivity
If a binary mask is used to protect pretrained weights, then computation is reduced, but the model may lose the ability to learn from all available data
Solution Approach 1:
The binary mask applies local quality by allowing different regions of the weight space to have different learnability characteristics. Updateable weights can learn from new data while non-updateable weights preserve pretrained knowledge, enabling the model to efficiently learn from available data without losing adaptability through selective activation of learning-capable parameters.
Data Source
AI summary
A learning method of an incremental learning model includes setting a searching range of mask weights based on either one or both of a distribution of mask weights of a binary mask corresponding to a filter of a pretrained model and a learning rate-related parameter, identifying a targeted mask weight in the searching range of the mask weights, updating the targeted mask weight based on the binary mask and the pretrained model, and updating a portion of the binary mask based on the updated targeted mask weight and a preset reference value.