By routing input tokens to only a few split experts, Sparse
Mixture-of-Experts has enabled efficient training of large language models.
Recent findings suggest that fixing the routers can achieve competitive
performance by alleviating the collapsing problem, where all experts eventually
learn similar representations. However, this strategy has two key limitations:
(i) the policy derived from random routers might be sub-optimal, and (ii) it
requires extensive resources during training and evaluation, leading to limited
efficiency gains. This work introduces \HyperRout, which dynamically generates
the router's parameters through a fixed hypernetwork and trainable embeddings
to achieve a balance between training the routers and freezing them to learn an
improved routing policy. Extensive experiments across a wide range of tasks
demonstrate the superior performance and efficiency gains of \HyperRouter
compared to existing routing methods. Our implementation is publicly available
at {\url{{this https URL}}}.

通过固定的超网络和可训练嵌入层动态生成路由器参数的方式，HyperRouter 在训练和冻结过程中学习优化路由策略，从而在各项任务上展现出明显的性能提升和效率增益。

HyperRouter: 面向稀疏专家混合的高效训练和推理

HyperRouter: Towards Efficient Training and Inference of Sparse Mixture  of Experts

Capsules as well as dynamic routing between them are most recently proposed
structures for deep neural networks. A capsule groups data into vectors or
matrices as poses rather than conventional scalars to represent specific
properties of target instance. Besides of pose, a capsule should be attached
with a probability (often denoted as activation) for its presence. The dynamic
routing helps capsules achieve more generalization capacity with many fewer
model parameters. However, the bottleneck that prevents widespread applications
of capsule is the expense of computation during routing. To address this
problem, we generalize existing routing methods within the framework of
weighted kernel density estimation, and propose two fast routing methods with
different optimization strategies. Our methods prompt the time efficiency of
routing by nearly 40\% with negligible performance degradation. By stacking a
hybrid of convolutional layers and capsule layers, we construct a network
architecture to handle inputs at a resolution of $64\times{64}$ pixels. The
proposed models achieve a parallel performance with other leading methods in
multiple benchmarks.

通过优化胶囊网络中的动态路由方法，提高路由效率，以实现与其他领先方法相媲美的性能表现。