Glossario IA
Il dizionario completo dell'Intelligenza Artificiale
Federated Learning
Distributed machine learning approach where models are trained on local data without centralizing it, thus preserving data privacy while collaborating on a global model.
Secure Aggregation
Cryptographic protocol allowing the combination of model updates from multiple clients in a secure manner, preventing the server from accessing individual client updates.
Gradient Compression
Technique for reducing the size of gradients transmitted between clients and server using methods like quantization or sampling to minimize bandwidth usage.
Adaptive Quantization
Compression method that dynamically adjusts the precision level of gradients based on their statistical importance, thus optimizing the trade-off between accuracy and communication volume.
Asynchronous Communication
Communication paradigm where clients can send their updates independently of others, reducing waiting times and improving the overall efficiency of the federated system.
Gradient Pruning
Technique consisting of transmitting only the most significant gradients by eliminating those with magnitude below a predefined threshold, thus significantly reducing network traffic.
Entropy Coding
Compression method that exploits the statistical properties of gradient distributions to encode information more efficiently, thus minimizing the size of transmitted data.
Client Selection Strategy
Algorithm optimizing the choice of participants in each training round based on factors such as connection quality, computing power, and relevance of local data.
Weighted Aggregation by Importance
Aggregation technique that assigns different weights to client updates based on the quality of their data and their contribution to improving the overall model.
Differentially Private Communication
Approach that adds controlled noise to communications to guarantee differential privacy, thus protecting individual information while maintaining the utility of the aggregated model.
Incremental Model Transmission
Strategy where only the differences between successive model versions are transmitted, significantly reducing the volume of data exchanged between training rounds.
Bandwidth Optimization
Set of techniques aimed at minimizing network bandwidth usage while maintaining model convergence, including compression, sampling, and intelligent scheduling.
Sparse Communication
Method consisting of transmitting only a fraction of model parameters or gradients, selected according to their importance, to drastically reduce communication volume.
Hierarchical Communication System
Network architecture organized in multiple levels where aggregation occurs progressively through intermediate nodes, thus reducing the load on the central server.
Adaptive Communication Protocol
Mechanism that dynamically adjusts the frequency and volume of communications based on network quality, model convergence, and available resources.
Preliminary Local Aggregation
Technique where clients perform several local optimization steps before communication, thus reducing the number of rounds needed to achieve convergence.
Model Compression via Knowledge Distillation
Method where a compact model learns to mimic the predictions of a larger model, thereby reducing the transmitted model size while preserving its performance.
Latency Optimization
Set of strategies aimed at minimizing communication delays in federated systems, including parallelism, network prediction, and intelligent scheduling.
Compressed Tensor Encoding
Advanced compression technique leveraging the tensor structure of gradients and weights to significantly reduce their transmission size without significant information loss.
Loss-Robust Communication
Protocol designed to tolerate packet loss and frequent disconnections in unstable network environments, ensuring convergence despite imperfect communications.