AI 용어집
인공지능 완전 사전
Self-Supervised Learning
Learning paradigm where the system generates its own labels from input data to create supervision signals without human intervention.
Contrastive Pre-training
Learning method that maximizes similarity between representations of positive instances and minimizes that of negative instances in latent space.
Masked Modeling
Technique where parts of the input data are masked and the model learns to reconstruct or predict these missing segments.
Reconstruction Learning
Approach where the network learns representations by training to reconstruct the original input from a degraded or encoded version.
Time Series
Ordered collection of data points indexed chronologically, exhibiting intrinsic temporal dependencies and correlations.
Temporal Embedding
Dense vector representation that captures dynamic features and temporal relationships of sequential data.
Long-Term Memory
Property of advanced architectures enabling the preservation and access to distant information in the temporal sequence.
Future Frame Prediction
Task specific to video sequences where the model generates or anticipates subsequent frames based on previous frames.
Spatio-Temporal Encoding
Process of transforming video data into vector representations simultaneously capturing spatial and temporal relationships.
Unsupervised Learning
Paradigm where the model learns relevant representations without using manual labels provided by human experts.
Pretext Task
Auxiliary learning objective designed to enable the acquisition of useful representations in the absence of direct supervision.
Dynamic Representations
Learned features that evolve and adapt to capture changes and transitions in temporal data.
Temporal Generative Model
Architecture capable of generating new realistic temporal sequences by learning the underlying data distribution.
Temporal Normalization
Preprocessing technique that standardizes temporal features to stabilize training and improve convergence.
Multi-Horizon Prediction
Model's ability to generate predictions for different future time intervals simultaneously from a single present state.