Enhanced Hybrid Hierarchical Federated Edge Learning Over Heterogeneous Networks

Chen, Qimei; You, Zehua; Wen, Dingzhu; Zhang, Zhaoyang

doi:10.1109/tvt.2023.3287355

Cited by 2 publications

(1 citation statement)

References 32 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…The cost of communicating is a big issue in global deep learning settings. Studies [13] on how to improve communication in distributed deep learning systems and [14] on how to train models efficiently in parallel show how important it is to fix issues with communication. But because NLP models are so different, they need specific ways to work together better.…”

Section: Review Of Literaturementioning

confidence: 99%

Optimizing Resource Allocation in Cloud for Large-Scale Deep Learning Models in Natural Language Processing

Gauri Dhopavkar

2024

jes

View full text Add to dashboard Cite

The need for big deep learning models in Natural Language Processing (NLP) keeps rising, it's important to find the best way to divide up cloud resources so that they can be used efficiently and at high speeds. This solves the problems that come with setting up and handling large NLP models by suggesting a complete strategy for making the best use of cloud-based platforms' resources. Combining model parallelism, data parallelism, and dynamic scaling methods, the suggested approach spreads the computing load across multiple cloud instances in better way. The framework constantly changes how resources are allocated to handle changes in workload by taking into account the specifics of NLP tasks, such as the need for different model designs and data processing needs. To improve scale and cut down on inference delay, a new auto-scaling method is introduced that lets computing resources be changed automatically based on demand in real time. The framework uses machine learning-based prediction models to figure out what resources will be needed in the future. This lets you make proactive decisions about scaling and keeps you from underusing or overprovisioning resources. It also solves the problem of communication overhead in distributed environments by improving data exchange protocols and using advanced inter-process communication techniques. The results of the experiments show that the proposed framework works well at improving both cost-effectiveness and prediction performance for large-scale NLP models by making the best use of resources. The framework is flexible enough to work with a wide range of natural language processing (NLP) tasks. It makes a useful addition to the efficient use of deep learning models in cloud settings.

show abstract

Section: Review Of Literaturementioning

confidence: 99%