Deep Metric Learning Using Triplet Network

By Elad Hoffer, Nir Ailon. 2014

2,210 citations recorded by the source corpus.

Deep learning has proven itself as a successful set of models for learning useful semantic representations of data. These, however, are mostly implicitly learned as part of a classification task. In this paper we propose the triplet network model, which aims to learn useful representations by distance comparisons. A similar model was defined by Wang et al. (2014), tailor made for learning a ranking for image information retrieval. Here we demonstrate using various datasets that our model learns a better representation than that of its immediate competitor, the Siamese network. We also discuss future possible usage as a framework for unsupervised learning.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, Radiology, Nuclear Medicine and Imaging, Representation & Self-Supervised Learning

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Computer Vision and Pattern Recognition

In Radiology, Nuclear Medicine and Imaging

In Representation & Self-Supervised Learning

Citations in the corpus

Matching Networks for One Shot Learning (2016, methodology): “Second, we employ a training strategy which is tailored for one-shot learning from the support set S.”
On Mutual Information Maximization for Representation Learning (2019, background): “…the structure present in the input data, namely spatial (e.g. patches from the same image should be closer than patches from different images) and temporal information (temporally close video frames should be encoded closer than the ones which are further away in time) (Hoffer and Ailon, 2015).”
Learning to See by Moving (2015, methodology): “Closest to our method is the the work of transforming auto-encoders [19] that used non-visual egomotion to reconstruct the transformed image from an input source image.”
Learning Deep Structure-Preserving Image-Text Embeddings (2015, background): “Finally, our work is related to deep similarity learning [4, 6, 7, 16, 18, 39, 47], though we are solving a cross-view, not a within-view, matching problem.”
Unsupervised Learning of Visual Representations Using Videos (2015, background): “Finally, our work is also related to metric learning via deep networks [51, 31, 4, 17, 15, 22, 54].”
Dynamic Few-Shot Visual Learning Without Forgetting (2018, methodology): “Moreover, our cosine-similarity based classification objective resembles the training objectives typically used by metric learning approaches [7].”
Few-Shot Adversarial Domain Adaptation (2017, methodology): “We address that problem by modifying the usual pairing technique used in many applications such as learning similarity metrics [7, 23, 22].”
Continual Learning through Retrieval and Imagination (2022, background): “Metric learning losses (Hoffer et al. 2015; He et al. 2021; Wang et al. 2019a) are used for training embedding networks, with the aim to reduce the distance between similar data and to increase the distance between dissimilar data.”
New Insights on Reducing Abrupt Representation Change in Online Continual Learning (2021, background): “Hou et al. (2019) considers addressing this imbalance through applying cosine similarity based loss as opposed to the typical cross entropy loss along with a distillation loss and a margin based loss with negatives mining to preserve the feature of previous classes.”
Semantic Drift Compensation for Class-Incremental Learning (2020, methodology): “The embedding network (trained with triplet loss [13]) represents classes with a prototype and performs NCM for classiﬁcation, and is denoted by E-FT.”
SphereFace: Deep Hypersphere Embedding for Face Recognition (2017, background): “Most widely used loss functions for deep metric learning are contrastive loss [1, 3] and triplet loss [32, 22, 6], and both impose Euclidean margin to features.”
Actions ~ Transformations (2015, methodology): “To model the transformation, we apply a Siamese network architecture in this paper, which is also related to the literature using deep metric learning [11, 9, 37, 12, 54, 16].”
Transitive Invariance for Self-Supervised Visual Representation Learning (2017, methodology): “In addition, the ranking function used in this paper is related to deep metric learning with Siamese architectures [5, 21, 19, 54, 24].”
Learning Two-Branch Neural Networks for Image-Text Matching Tasks (2017, background): “Our two-branch networks are related to Siamese networks for metric learning [44], [45], [46], [47], [48], [49].”
Improving Generalization via Scalable Neighborhood Component Analysis (2018, methodology): “ subsequent linear classier. Metric learning learning encourages the minimization of intra-class variations and the maximization inter-class variations, such as contrastive loss [1,37], triplet loss [14]. Recent works on few-shot learning [42,36] also show the utility of metric learning, since it is”
Variational Prototype Replays for Continual Learning (2019, methodology): “ (Section 4.2). In contrast with them, our method compares the feature similarities represented by class prototypes in the embedding space which improves generalization, as also been veriﬁed in works [15, 32]. Intuitively, if the number of data samples in D~ tis very large, the network could re-prod”
Learning Metrics From Teachers: Compact Networks for Image Embedding (2019, methodology): “Later Triplet networks were proposed which allow more local modifications of the embedding space, and do not require that all the observations of the same class collapse to the same point [12, 35].”
Unsupervised Learning of Semantic Audio Representations (2017, background): “We instead resort to deep metric learning using triplet loss [7, 8, 9], which finds a nonlinear mapping into a low dimensional space where simple Euclidean distance can express any desired relationship between examples of the form X is more like Y than like Z.”
CosFace: Large Margin Cosine Loss for Deep Face Recognition (2018, methodology): “[5, 7, 10, 39] require thoroughly scheming the mining of pair or triplet samples, which is an extremely time-consuming procedure.”