- Constellation Loss: Improving the ef铿乧iency of deep metric learning loss functions for optimal embedding. [NIPS] Alfonso Medela, Artzai Picon.
- when we use the multi-class n pair loss, the num of negative is fixed. because of using the the same class to be the negative sample.
- And the controlled parameters K, means we can change the computational effort in the model.
- FaceNet: A Uni铿乪d Embedding for Face Recognition and Clustering [CVPR 2015] Florian Schroff, Dmitry Kalenichenko, James Philbin.
- a unified embedding for Recognition and Clustering
- Triple Loss to align the distance between Position sample and negative sample
- online negative sample mining strategy
- Beyond triplet loss: a deep quadruplet network for person re-identification [CVPR 2017] Weihua Chen, Xiaotang Chen, Jianguo Zhang, Kaiqi Huang.
- large inter-class variation + small intra-class variation
- $$\begin{aligned} L_{q u a d}=& \sum_{i, j, k}^{N}\left[g\left(x_{i}, x_{j}\right)^{2}-g\left(x_{i}, x_{k}\right)^{2}+\alpha_{1}\right]{+} \ &+\sum{i, j, k, l}^{N}\left[g\left(x_{i}, x_{j}\right)^{2}-g\left(x_{l}, x_{k}\right)^{2}+\alpha_{2}\right]{+} \ & s{i}=s_{j}, s_{l} \neq s_{k}, s_{i} \neq s_{l}, s_{i} \neq s_{k} \end{aligned}$$
- Improved Deep Metric Learning with Multi-class N-pair Loss Objective [CVPR 2016] Kihyuk Sohn.
- $$\mathcal{L}{N-\text { pair me }}\left(\left{\left(x{i}, x_{i}^{+}\right)\right}{i=1}^{N} ; f\right)=\frac{1}{N} \sum{i=1}^{N} \log \left(1+\sum_{j \neq i} \exp \left(f_{i}^{\top} f_{j}^{+}-f_{i}^{\top} f_{i}^{+}\right)\right)$$
- Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks [EMNLP 2019] Nils Reimers, Iryna Gurevych.
- reduce similar computational in STS.
- three pooling strategy
- CLS-token
- max strategy
- mean strategy
- three objective function
- classification objective function
- regression objective function
- triplet objective function
Repository navigation
MetricLearning
Directory actions
More options
Directory actions
More options
聽
聽