Cosine Classifier — A Direction-Based Classifier Driven by Angular Similarity

Cosine Classifier is a classification approach that uses cosine similarity between an input feature vector and each class’s weight vector. Unlike a standard linear classifier that mixes both magnitude and direction, a cosine classifier focuses on the direction (angle) of vectors. This makes it especially useful when the angular structure of representations matters more than raw feature scale.

03/05/2026

Feature Embedding — A Vector Representation That Maps Data into a Semantic Coordinate Space to Capture Similarity and Structure

Feature Embedding is a representation method that converts input data such as images or text into fixed-length vectors that encode semantic similarity and underlying structure. These vectors are not just numbers; they act as coordinates in a space where relationships between data points can be expressed through distance and direction.

03/20/2026

Fine-grained Classification — Learning the Subtle Features That Separate Similar Classes

Fine-grained Classification focuses on distinguishing subtle visual differences between classes that appear almost identical. Unlike conventional classification tasks that separate broad categories such as animals, vehicles, or objects, Fine-grained Classification deals with differences within the same category, such as identifying bird species or distinguishing between similar car models. The main challenge is not recognizing the overall shape of an object, but learning the local features and precise representation patterns that define each class.

07/23/2026

Metric Learning — Learning Distance Between Data Points to Represent Similarity

Metric Learning is a way of learning a distance metric that represents similarity between data points. Instead of only learning a decision boundary for classification, it learns an embedding space where similar samples are pulled closer together and dissimilar samples are pushed farther apart. This makes it especially useful in tasks where “how similar are these two items?” matters more than assigning a fixed class label, such as face recognition, image retrieval, and recommendation systems.

03/09/2026

Triplet Loss (Triplet Loss) — Distance-Based Loss that Learns to Make Similar Samples Closer and Dissimilar Samples Far Apart Beyond a Margin

Triplet Loss (Triplet Loss) is a distance-based loss function widely used in representation learning (Representation Learning). Its goal is simple: make samples from the same class closer in the embedding space while pushing samples from different classes farther apart. In other words, instead of simply learning to “predict the correct class,” the model learns the similarity structure between data points. This approach is particularly powerful in tasks such as face recognition, image retrieval, and recommendation systems, where “similar items should cluster and dissimilar ones should separate.”

03/09/2026