Metric learning by discriminant neighborhood embedding

Hebbian semi-supervised learning in a sample efficiency setting

“…Moreover, bio-inspired extensions of max-margin approaches for local learning have been proposed [55,61,162]. It is worth mentioning that the idea of augmenting unsupervised methods with label information has been explored in discriminative clustering [113,125], discriminative subspace learning [14,17,148,200], discriminative sparse coding [159,160,243], discriminative ICA [4,30,58], and discriminative manifold learning [40,54,76,150,196,235,249].…”

Scalable Similarity Learning Using Large Margin Neighborhood Embedding

“…We still need to define the weights w ij , which encode the class label information. Binary values ±1 are commonly used to assign w ij based on whether x i and x j come from the same class or not [23], which essentially has the same effect as partitioning the adjacency graph into a within-class graph and a between-class graph [22,29,30] to impose a pairwise constraint that the similarity should be high between samples from the same class and low otherwise. For kNN classifiers, as suggested in [11], we care more about the relative similarity defined over a triplet of samples; i.e., a higher similarity score should be assigned when a sample x i is compared with any of its target neighbor x j , j∈N + i ∨ i∈N + j than with any of its imposter neighbor x l , l∈N − i ∨ i∈N − l .…”

“…In this paper, we propose a new similarity learning algorithm that features good scalability with respect to both sample size and dimensionality. First, motivated by the findings from manifold learning with neighborhood embedding [22,23], we restrict similarity comparison to sample pairs within the same local neighborhood, and try to capture the discriminative structure of local data manifold using large margin neighborhood embedding (Sec. 2).…”

Local discriminative distance metrics ensemble learning

“…Distance metric learning plays a crucial role in metric-related pattern recognition tasks including K-means, K-Nearest Neighbors, and kernel-based algorithms such as SVMs [19,4,5,23,25]. The learning task falls into two categories: unsupervised and supervised distance metric learning.…”