AI 용어집
인공지능 완전 사전
Kubernetes for ML
Kubernetes container orchestration adapted for machine learning workloads, including GPU management, horizontal scaling of distributed training, and automated deployment of inference models.
GPU Clustering
Aggregation of multiple GPUs into a unified computational cluster enabling data and model parallelism to accelerate large-scale deep neural network training.
Distributed Training
ML model training technique distributing the computational load across multiple nodes, using strategies like data parallelism or model parallelism to reduce convergence time.
Resource Pooling
Virtualization and dynamic sharing of computational resources (CPU, GPU, memory) between different ML tasks, optimizing utilization and reducing infrastructure costs.
Autoscaling ML
Mechanism for automatic adaptation of computational resources based on ML workload metrics, ensuring optimal performance during training or inference peaks.
Container Orchestration
Automation of deployment, scaling, and management of ML application containers, including service discovery, load balancing, and resilience against failures.
Inference Optimization
Set of techniques (quantization, pruning, distillation) aimed at reducing model latency and memory consumption during the production inference phase.
Real-time Inference
Infrastructure capable of providing predictions with minimal latency (generally <100ms), essential for critical applications like fraud detection or recommendation systems.
Edge Computing ML
Deployment of ML models on edge devices to reduce latency, preserve data privacy, and minimize dependency on network connectivity.
Cloud Native ML
Architectural approach leveraging native cloud services for the complete ML lifecycle, from distributed training to serverless model deployment.
Model Versioning Infrastructure
ML model versioning system with artifact tracking, training metadata, and rollback capabilities to ensure traceability and reproducibility.
Load Balancing ML
Intelligent distribution of inference requests across multiple model instances, based on CPU/GPU load and prediction complexity to optimize response times.
Cluster Management
Monitoring and administration of computational node sets for ML, including provisioning, monitoring, and maintenance of training and inference clusters.
Spot Instance Management
Strategy for using low-cost cloud spot instances for non-critical ML workloads, with checkpointing and migration mechanisms to handle interruptions.
GPU Scheduling
Optimized allocation and scheduling of ML tasks on available GPU resources, maximizing throughput while respecting job priorities and constraints.
Multi-Cloud ML Deployment
ML model deployment strategy across multiple cloud providers for redundancy, cost optimization, and regulatory data compliance.
Serverless ML
ML architecture without explicit server management, where infrastructure automatically scales to load, billed only for actual resource usage.
Infrastructure as Code for ML
Automation of ML infrastructure provisioning and configuration via declarative code, ensuring reproducibility and versioned management of environments.