No abstract

Training multiple deep neural networks (DNNs) and averaging their outputs is a simple way to improve the predictive performance. Nevertheless, the multiplied training cost prevents this ensemble method to be practical and efficient. Several recent works attempt to save and ensemble the checkpoints of DNNs, which only requires the same computational cost as training a single network. However, these methods suffer from either marginal accuracy improvements due to the low diversity of checkpoints or high risk of divergence due to the cyclical learning rates they adopted. In this paper, we propose a novel method to ensemble the checkpoints, where a boosting scheme is utilized to accelerate model convergence and maximize the checkpoint diversity. We theoretically prove that it converges by reducing exponential loss. The empirical evaluation also indicates our proposed ensemble outperforms single model and existing ensembles in terms of accuracy and efficiency. With the same training budget, our method achieves 4.16% lower error on Cifar-100 and 6.96% on Tiny-ImageNet with ResNet-110 architecture. Moreover, the adaptive sample weights in our method make it an effective solution to address the imbalanced class distribution. In the experiments, it yields up to 5.02% higher accuracy over single EfficientNet-B0 on the imbalanced datasets.

Federated learning (FL) is a privacy-preserving paradigm for multi-institutional collaborations, where the aggregation is an essential procedure after training on the local datasets. Conventional aggregation algorithms often apply a weighted averaging of the updates generated from distributed machines to update the global model. However, while the data distributions are non-IID, the large discrepancy between the local updates might lead to a poor averaged result and a lower convergence speed, i.e., more iterations required to achieve a certain performance. To solve this problem, this paper proposes a novel method named AggEnhance for enhancing the aggregation, where we synthesize a group of reliable samples from the local models and tune the aggregated result on them. These samples, named class interior points (CIPs) in this work, bound the relevant decision boundaries that ensure the performance of aggregated result. To the best of our knowledge, this is the first work to explicitly design an enhancing method for the aggregation in prevailing FL pipelines. A series of experiments on real data demonstrate that our method has noticeable improvements of the convergence in non-IID scenarios. In particular, our approach reduces the iterations by 31.87% on average for the CIFAR10 dataset and 43.90% for the PASCAL VOC dataset. Since our method does not modify other procedures of FL pipelines, it is easy to apply to most existing FL frameworks. Furthermore, it does not require additional data transmitted from the local clients to the global server, thus holding the same security level as the original FL algorithms.