MoDE: A Mixture-of-Experts Model with Mutual Distillation among the Experts

The application of mixture-of-experts (MoE) is gaining popularity due to its ability to improve model's performance. In an MoE structure, the gate layer plays a significant role in distinguishing and routing input features to different experts. This enables each expert to specialize in processing their corresponding sub-tasks. However, the gate's routing mechanism also gives rise to narrow vision: the individual MoE's expert fails to use more samples in learning the allocated sub-task, which in turn limits the MoE to further …

application cs.ai cs.lg distillation expert experts features gate layer moe performance processing role routing s performance tasks