AI 용어집
인공지능 완전 사전
Inference Engine
Optimized software that executes pre-trained AI models on target hardware, managing resource allocation, computation scheduling, and leveraging hardware accelerations for real-time inference.
Pruning
Model compression technique that selectively removes the least influential weights or neurons from a neural network, thereby reducing its computational complexity for edge deployment.
Neural Architecture Search (NAS) for Edge
Process of automating the design of AI models specifically optimized for the hardware constraints of edge devices, balancing model accuracy with its size, latency, and energy consumption.
Low-Power System-on-Chip (LP-SoC)
Integrated circuit that combines a CPU, hardware accelerators for AI (NPU), memory, and interfaces on a single chip, designed for minimal energy consumption while providing HPC performance suitable for edge computing.
Real-Time Operating System (RTOS) for AI
Lightweight and deterministic operating system, optimized for running AI workloads with strict latency guarantees, essential for critical applications like autonomous driving or robotic control at the edge.
Continual Learning on Device
Ability of an AI model to continuously adapt and learn from new data directly on the edge device, without forgetting previous knowledge and without requiring centralized retraining.
Sparse Computing
Computing paradigm that exploits zeros (sparsity) in AI model weights and activations to skip unnecessary operations, thereby drastically reducing the number of computations and energy consumption on edge hardware.
Secure Enclave for AI
Isolated hardware security zone within a processor, designed to execute sensitive AI workloads confidentially, protecting models and data from unauthorized access even if the main system is compromised.