AI 용어집
인공지능 완전 사전
Few-Shot Object Detection
A sub-discipline of object detection aimed at training models capable of recognizing new object categories using a very limited number of training examples, often fewer than ten per class.
Metric Learning
A technique that involves learning a distance function or embedding space where objects of the same class are close and those of different classes are far apart, facilitating classification with few examples.
Prototype Network
A meta-learning model that builds a 'prototype' for each class by averaging the features of the few available examples, then classifies a new instance by matching it to the nearest prototype.
Matching Network
An architecture that uses an attention mechanism to compare a new example (query) to the set of support examples, generating a weighted prediction based on the similarity of each support example.
Relation Network
An extension of matching networks that explicitly learns a similarity function (relation module) to measure the degree of relation between a query and the support examples, instead of using a fixed distance.
Support-Based Fine-Tuning
A strategy where a model pre-trained on a large base dataset is quickly re-trained (fine-tuned) on the very small support dataset (few-shot) to adapt to the new classes.
Embedding Space
A low-dimensional vector representation where objects are projected, designed so that the distance between vectors reflects the semantic or visual similarity of the objects, which is crucial for classification with few examples.
Contrastive Loss
A loss function used in metric learning that minimizes the distance between pairs of examples from the same class (anchors and positives) while maximizing the distance to examples from different classes (negatives).
Triplet Loss
Loss function that operates on triplets of examples (an anchor, a positive of the same class, a negative of a different class) to learn an embedding space where the anchor-positive distance is smaller than the anchor-negative distance by a certain margin.
Box-Free Detection
Alternative paradigm in few-shot that focuses on classifying the presence of objects in a region without predicting precise bounding boxes, thus simplifying the task when data is extremely limited.
Reinforcement Learning for Few-Shot
Use of reinforcement learning algorithms to optimize the few-shot learning process, for example, by learning a policy to select the most relevant support examples or to adjust the model.
Class Re-weighting
Technique that adjusts the contribution of each class in the loss function to counter the extreme imbalance between base classes (numerous) and new classes (rare) in a few-shot detection context.
One-Shot Object Detection
Extreme case of few-shot where the model must learn to detect a new object class from a single labeled example, representing a major challenge in terms of generalization.
Zero-Shot Object Detection
An even more ambitious scenario than few-shot, where the model must detect object classes for which it has received no visual examples, relying solely on semantic descriptions (e.g., attributes or text).