Knowledge Graph — Coursera Notes › Academic disciplines › Information Technology / Computer Science › Machine Learning

Federated Learning

concept · part of Machine Learning

Federated learning is a machine learning approach that trains models across decentralized devices or locations without centralizing raw data. It combines decentralized computation with privacy measures, enabling collaborative model training while keeping sensitive data local. The process involves a global model sent to participating devices, local training using on-device data, transmission of model updates (not raw data) to a central server, aggregation of updates (e.g., via federated averaging), and redistribution of the refined global model. This iterative cycle improves model performance without exposing private data. Key advantages include enhanced privacy and security, personalized training, and bandwidth efficiency. Challenges include data variability (non-IID data), secure communication, and resource limitations on devices.

Federated averaging combines individual updates (which may be noisy or biased due to non-IID data) to refine the global model, ensuring convergence while adhering to privacy constraints.

• Training topology: Decentralized; model weights and gradients are exchanged instead of raw data.
• Central server role: Initializes a shared model and synchronizes updates from edge devices.
• Local optimization: Each device performs local optimization (e.g., stochastic gradient descent) to generate model parameters.

Federated learning preserves privacy by keeping raw data on local devices and only transmitting model updates. This reduces risks of data breaches and unauthorized access, and ensures compliance with privacy regulations. Bandwidth is also reduced since only model parameters are transmitted, not entire datasets.

Federated learning is applied in various industries:

• Health care: Hospitals collaboratively train models for disease prediction or diagnostics without sharing sensitive patient data.
• Smartphones: Google uses federated learning to improve predictive text and keyboard suggestions by training models directly on users' devices.
• IoT and smart homes: Devices like smart thermostats or cameras collaboratively train models to optimize energy usage or improve security systems.

• Health care diagnostics: Traditional ML centralizes patient data on a server risking breaches, while federated learning keeps data on hospital servers or patient devices.
• Smartphone personalization: Traditional ML collects user typing data centrally, whereas federated learning trains models on-device.
• IoT optimization: Traditional ML sends raw data to a cloud server increasing bandwidth, while federated learning transmits only model updates.

Inside Federated Learning (2)

• Non-IID Data in Federated Learning — Non-IID data, where each device's local data distribution differs from the global distribution, poses a significant challenge in federated learning.
• TensorFlow Federated — Federated learning can be implemented using frameworks such as TensorFlow Federated and PyTorch.

This is the text view of an interactive 3D knowledge graph — open this page with JavaScript enabled to explore it visually.