Glosarium AI
Kamus lengkap Kecerdasan Buatan
Rotation Prediction
Self-supervised learning task where the model predicts the rotation angle applied to an image, enabling the learning of visual representations without manual annotation.
Proxy Task
Intermediate learning objective designed to pre-train models on unlabeled data, serving as a substitute for the final target task.
Self-Supervision
Learning paradigm where labels are automatically generated from the input data, eliminating the need for human annotations.
Feature Learning
Process of automatically extracting discriminative features from raw data, without manual feature engineering.
Rotational Invariance
Property of a visual representation that remains stable despite rotations of the input image, essential for model robustness.
Visual Representation
Compact vector encoding of semantic information contained in an image, learned by deep neural networks.
Rotation Consistency
Principle according to which features extracted from an image and its rotated versions should share common semantic properties.
Rotation Augmentation
Data augmentation technique applying random rotations to images to create diversity and improve model generalization.
Rotation Prediction Loss
Loss function measuring the discrepancy between the predicted rotation angle and the actual angle, guiding the learning of visual representations.
Visual Feature Space
High-dimensional vector space where similar images are projected close to each other after self-supervised learning.
Rotation Classification
Formulation of rotation prediction as a discrete classification problem (0°, 90°, 180°, 270°) rather than continuous regression.
Self-Supervised Pretraining
Initial training phase using automatically generated supervised signals, before fine-tuning on the target task.
Rotation-Agnostic Features
Visual features that capture semantic content independently of the spatial orientation of the object in the image.
Geometric Transformation
Family of spatial transformations including rotations, translations, and flips used as proxy tasks in self-supervised learning.
Rotation Embedding
Learned vector representation that implicitly encodes spatial orientation information of objects in the latent space.
Contrastive Rotation Learning
Approach combining rotation prediction with contrastive learning to strengthen the separation between different orientations.