CAMERO: Consistency Regularized Ensemble of Perturbed Language Models with Weight Sharing
This addresses the deployment challenge of ensemble models in real-world applications by reducing memory costs while maintaining performance, though it is incremental as it builds on existing weight-sharing and perturbation techniques.
The paper tackles the high memory and inference costs of model ensembles by proposing CAMERO, a consistency-regularized ensemble with weight sharing and perturbations, which improves generalization and reduces model size, outperforming an 8-model BERT-base ensemble on GLUE by 0.7 points with a model size of 114.2M versus 880.6M.
Model ensemble is a popular approach to produce a low-variance and well-generalized model. However, it induces large memory and inference costs, which are often not affordable for real-world deployment. Existing work has resorted to sharing weights among models. However, when increasing the proportion of the shared weights, the resulting models tend to be similar, and the benefits of using model ensemble diminish. To retain ensemble benefits while maintaining a low memory cost, we propose a consistency-regularized ensemble learning approach based on perturbed models, named CAMERO. Specifically, we share the weights of bottom layers across all models and apply different perturbations to the hidden representations for different models, which can effectively promote the model diversity. Meanwhile, we apply a prediction consistency regularizer across the perturbed models to control the variance due to the model diversity. Our experiments using large language models demonstrate that CAMERO significantly improves the generalization performance of the ensemble model. Specifically, CAMERO outperforms the standard ensemble of 8 BERT-base models on the GLUE benchmark by 0.7 with a significantly smaller model size (114.2M vs. 880.6M).