Nearest Class Mean Metric Learning
2 likes440 views
The document discusses metric learning for image classification, particularly focusing on Nearest Class Mean Metric Learning (NCMML) and its expansion to Nearest Class Multiple Centroids (NCMC). NCMML allows for flexible classification by minimizing distances between images and class prototypes using metrics, and it shows performance improvements over traditional SVMs. The document also mentions the implementation considerations and available resources for coding these methods in MATLAB.
Nearest Class Mean Metric Learning
- 1. T. Mensink et al., Distance-Based Image Classification: Generalizing to new classes at near-zero cost, IEEE PAMI, April, 2013 Reviewed by LG CNS, AI Tech Team Sangjun Han 2018. 12. 09
- 2. 2 Metric Learning • What is metric learning - Learning metrics for solving machine learning problems - Simply speaking, a loss function can be defined using metrics - e.g. Euclidean distance, cosine similarity… • Why metric learning - General classification problems => softmax(losses for each class) - Their output values tell us which is more probable, but no more information - Metric learning => not only single probability, but also metrics between its classes - e.g. The distance (or similarity) between class 1 and class 2 - Image retrieval, face verification, incremental learning… more flexible!
- 3. 3 Nearest Class Mean Metric Learning (NCMML) • Concept - Assign an x (image or feature from image) to the class where it is closest to class prototype uc - The prototype can be obtained by averaging input training data in a class (So, you can get C prototypes which stands for each class) - Mahalanobis distance between x and uc We learn W to minimize distance of x and x’ It means to minimize Euclidean distance in semantic space
- 4. 4 Nearest Class Mean Metric Learning (NCMML) Wu1 Wu2 Wu3 Wx
- 5. 5 Nearest Class Mean Metric Learning (NCMML) • Concept - Softmax activation for distance matrix - Cross entropy loss function - Gradient descent for the loss function
- 6. 6 Nearest Class Mean Metric Learning (NCMML) W : d x D X : D x N Mmeans : D x C (prototype for each class) D : feature dimension in original space d : feature dimension in sematic space N : The number of data C : The number of class Forward pass How to compute Euclidean distance efficiently between two matrices https://scikit-learn.org/stable/modules/generated/sklearn.metrics.pairwise.euclidean_distances.html • Implementation - You can get the author’s code at (written on MATLAB) https://staff.fnwi.uva.nl/t.e.j.mensink/uva12/code.php
- 7. 7 Nearest Class Mean Metric Learning (NCMML) Compute gradient • Implementation - You can get the author’s code at (written on MATLAB) https://staff.fnwi.uva.nl/t.e.j.mensink/uva12/code.php
- 8. 8 Nearest Class Mean Metric Learning (NCMML) • Results - NCMML performs well comparable to a SVM baseline (one-vs-rest linear SVM) - 8,500 fold speed-up faster than the SVM baseline - Generalized well to unseen classes at a negligible cost
- 9. 9 Nearest Class Multiple Centroids (NCMC) • Expansion of NCMML - Multiple prototypes (or centroids) for each class => non-linear classification - The prototypes can be obtained by k-means clustering (k is parameter) when k = 3
- 10. 10 Etc • No available online code in Python - PyDML v0.0.1 (by Juan Luis) supports NCMML and NCMC - However, NCMML is not working on the small machine (it requires hard computation for matrix outer product, not efficient code) - In NCMC, wrongly written from the paper - Recommend to write your code based on the author’s Matlab code Sangjun Han hjun1008@gmail.com https://www.linkedin.com/in/sangjun-han-78a166b8/