AI Briefing[CVPR 2026] An Upcycling Method for Creating True Experts within MoE Architecture - LG AI Research BLOG
Key point
LG AI Research has proposed 'Cluster-Aware Upcycling,' a technology that solves the problem of redundancy among experts when converting existing Dense models into MoE.
Details
The approach of increasing model size to improve AI model performance leads to massive computational costs. To address this, the Mixture-of-Experts (MoE) architecture, which activates only necessary parts, is gaining attention, but the existing Sparse Upcycling method simply copies the parameters of a pretrained Dense model as-is, causing an 'Expert Symmetry' problem where experts end up performing similar functions to one another.
To solve this, LG AI Research proposed Cluster-Aware Upcycling technology. This method is based on the insight that the activation values of an already well-trained Dense model contain the semantic structure of the data.
The core process is as follows:
- FFN Activation Value Clustering: After extracting the activation values input into the FFN block of the pretrained model, Spherical k-means is used to cluster semantically similar data regions.
- Expert-Specific Initialization: Based on the clustering results, different initial values are set so that each Expert takes charge of a different data region.
Through this technology, the model is guided from the initial stage so that each expert learns a unique region, enabling it to secure much more efficient and powerful specialization compared to existing methods.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
Sakana AI Releases 'Augmented Lagrangian Predictive Coding' Capable of Training 1000-Layer Neural Networks