Deep Learning baixar o livro de graça

Deep learning models typically require data to be collected at a centralized location to learn effective representations which has several implicit issues like communication cost, risk of data privacy, etc. These issues are even more severe in the case of clinical data, where patient privacy has the utmost priority. In such a context, distributed machine learning provides a way forward where various data holding sites can locally train a mutually decided model and share their knowledge. Federated learning (FL) provides such a mechanism with the help of a client-server framework. Clients in the FL environment are independent small edge devices that hold their data locally. In contrast, the server is a central site that aggregates and distributes the knowledge learned by each client to others. The server receives the locally trained weights from all the participating clients to perform aggregation of weights. Server then transfers the aggregated weights to all clients before starting of next round of training. This procedure repeats until the server attains the desired accuracy. Therefore FL enables multiple clients to collaboratively train a shared global model without sharing their local data. Fl learning not only preserve the data privacy but also overcomes the issue of limited data availability.
However FL itself suffers from various challenges such as high communication costs to transfer weights, statistical data heterogeneity among clients and single point failure of server. Client heterogeneity occurs mainly due to difference of data distribution at different clients and their associated computation power. This project targets statistic data heterogeneity in the FL environment and proposed a simple yet effective attention-based approach to handle the same. Precisely, in the proposed setting, each client sends a mean representation to the centralised server along with the trained model’s weights. A similarity matrix is computed based on similarity score of a clients’ using mean representation from every other participating client. The similarity matrix decide the weightage of client’s mode in the aggregated model. Specifically the centralised server computes the attention vector for each client using this similarity matrix and then broadcast this attention vector to all clients. We perform this attention mechanism on both the centralized server as well as the participating clients. We consider FedAvg, FedProx and FedMomentum as a baseline to be compared and our proposed approach outperforms all of them. For the statistical heterogeneity, we perform extensive experiments on FOOD101 and CIFAR10 which show that our approach performs well in the case of highly skewed data. Further to address the issue of single point failure in the FL we propose an efficient version of swarm learning. Swarm learning is a server less decentralized machine learning in its conventional form. We propose context-aware hierarchical swarm learning which is able to tackle the issue of both data heterogeneity as well as single point failure. We showcase the effectiveness of context-aware swarm learning we perform set of experiments on HAM10000 and ISIC Skin Lesion 2019 datasets. To address the issue of high communication cost in FL we propose BAFL (Federated Learning for Base Ablation) where we introduce a fine-tuning approach to leverage the feature extraction ability of layers at different depths of deep neural networks. We evaluate the proposed approach using VGG-16 and ResNet-50 models on datasets including WBC, FOOD-101, and CIFAR-10 and obtain up to two orders of reduction in total communication cost as compared to the conventional federated learning.

Procurando Deep Learning? Aqui você encontra tudo sobre este livro de Shweta Sankhwar. Nesta página estão a descrição da obra, os detalhes da edição (40 páginas) e os formatos disponíveis para baixar: pdf, mp3, áudio-livro, kindle. Se você gosta de Livros Internacionais, Engenharia e Transporte, Engenharia, Tecnologia da Computação, Sensoriamento Remoto, explore também outros títulos da mesma categoria no LivrosQualidade. Veja ainda as outras obras de Shweta Sankhwar em nosso catálogo.