Towards a Universal Gating Network for Mixtures of Experts
This work addresses the challenge of integrating diverse expert networks without access to training data, which is incremental as it builds on existing mixtures of experts by extending to heterogeneous and data-free scenarios.
The paper tackles the problem of combining heterogeneous pre-trained neural networks in a data-free regime, proposing multiple methods including specialized gating networks, and finds that gating networks achieve the highest accuracy.
The combination and aggregation of knowledge from multiple neural networks can be commonly seen in the form of mixtures of experts. However, such combinations are usually done using networks trained on the same tasks, with little mention of the combination of heterogeneous pre-trained networks, especially in the data-free regime. This paper proposes multiple data-free methods for the combination of heterogeneous neural networks, ranging from the utilization of simple output logit statistics, to training specialized gating networks. The gating networks decide whether specific inputs belong to specific networks based on the nature of the expert activations generated. The experiments revealed that the gating networks, including the universal gating approach, constituted the most accurate approach, and therefore represent a pragmatic step towards applications with heterogeneous mixtures of experts in a data-free regime. The code for this project is hosted on github at https://github.com/cwkang1998/network-merging.