Dynamic Mixture of Experts for Adaptive Computation in Character-Level Transformers
This paper challenges the prevailing assumption that Mixture of Experts (MoE) consistently improves computational efficiency through a systematic evaluation of MoE variants in Transformer models. We implement and compare three approaches: basic MoE, top-<i>k</i> routing, and capacity-fac...
Saved in:
| Main Authors: | , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
MDPI AG
2025-06-01
|
| Series: | Information |
| Subjects: | |
| Online Access: | https://www.mdpi.com/2078-2489/16/6/483 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|