Scribd
Upload
112 views15 pages

CS-671: Deep Learning and Its Applications Distance Metric Learning

This document summarizes a lecture on distance metric learning and siamese networks. It discusses how siamese networks can be used to learn a similarity metric for matching pairs of images by using a shared CNN to encode each image, calculating the distance between encodings, and optimizing a contrastive loss on positive and negative image pairs. Specifically, it uses Euclidean distance for positive pairs and hinge loss for negative pairs to learn image embeddings that bring similar images together and separate different images by a margin. The siamese network architecture can be varied for different tasks based on empirical performance.

Uploaded by

Shashi Kumar Fagna
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
112 views15 pages

CS-671: Deep Learning and Its Applications Distance Metric Learning

This document summarizes a lecture on distance metric learning and siamese networks. It discusses how siamese networks can be used to learn a similarity metric for matching pairs of images by using a shared CNN to encode each image, calculating the distance between encodings, and optimizing a contrastive loss on positive and negative image pairs. Specifically, it uses Euclidean distance for positive pairs and hinge loss for negative pairs to learn image embeddings that bring similar images together and separate different images by a margin. The siamese network architecture can be varied for different tasks based on empirical performance.

Uploaded by

Shashi Kumar Fagna
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 15

CS-671: Deep Learning and its Applications

Lecture: 16
Distance metric learning

Aditya Nigam, Assistant Professor


School of Computing and Electrical Engineering (SCEE)
Indian Institute of Technology, Mandi
http://faculty.iitmandi.ac.in/ãditya/ [email protected]

Presentation for CS-671@IIT Mandi (16 April, 2019)


(*Presented by : Daksh Thapar)

February - May, 2019


Need for Similarity Measures

Several applications of Similarity Measures exist in today’s world.


Handwriting recognition.

Detection of faces in camera image.

Aditya Nigam (SCEE, IIT-Mandi) [email protected] Lecture, February - May, 2019


Notion of a Metric

A Metric is a function that quantifies a “distance00 between every


pair of elements in a set, thus inducing a measure of similarity.

A metric f (x, y ) must satisfy the following properties for all x, y , z


belonging to the set:
Non-negativity: f (x, y )0
Identity of Discernible: f (x, y ) = 0 <=> x = y
Symmetry: f (x, y ) = f (y , x)
Triangle Inequality: f (x, z) <= f (x, y ) + f (y , z)

Aditya Nigam (SCEE, IIT-Mandi) [email protected] Lecture, February - May, 2019


Traditional Approaches for Matching

The traditional approach for matching images, relies on the following


pipeline:
1 Extract Features: For instance, color histograms of the input
images.
2 Learn Similarity: Use L1 − norm on the features.

Challenges:
The principal shortcoming of traditional metric learning based methods is
that the feature representation of the data and the metric are not learned
jointly.

Aditya Nigam (SCEE, IIT-Mandi) [email protected] Lecture, February - May, 2019


Problem Statement

Input: Given a pair of input images, we want to know how “similar”


they are to each other.

Output: The output can take a variety of forms:


Either a binary label, i.e. 0 (same) or 1 (different).
A Real number indicating how similar a pair of images are.

Aditya Nigam (SCEE, IIT-Mandi) [email protected] Lecture, February - May, 2019


Siamese Networks

Aditya Nigam (SCEE, IIT-Mandi) [email protected] Lecture, February - May, 2019


Siamese Networks as binary classification

Aditya Nigam (SCEE, IIT-Mandi) [email protected] Lecture, February - May, 2019


Problem Statement

Input: Given a pair of input images, we want to know how “similar”


they are to each other.

Output: The output can take a variety of forms:


Either a binary label, i.e. 0 (same) or 1 (different).
A Real number indicating how similar a pair of images are.
d(img 1, img 2) = Degree of difference between images
1 d(img 1, img 2) < t For images belonging to same class
2 d(img 1, img 2) > t For images belonging to different class

Aditya Nigam (SCEE, IIT-Mandi) [email protected] Lecture, February - May, 2019


Standard Siamese CNN

Input: A pair of input images.


Output: A label, 0 for similar, 1 else.

Aditya Nigam (SCEE, IIT-Mandi) [email protected] Lecture, February - May, 2019


Siamese CNN - Loss Function

Is there a problem with this formulation?

Aditya Nigam (SCEE, IIT-Mandi) [email protected] Lecture, February - May, 2019


Siamese CNN - Loss Function

Yes.
The model could learn to embed every input to a same point i.e.
predict a constant as output.
In such a case, every pair of input would be categorized as positive pair.
Aditya Nigam (SCEE, IIT-Mandi) [email protected] Lecture, February - May, 2019
Siamese CNN - Loss Function

The final loss is defined as:

L = Σloss of positive pairs + Σloss of negative pairs

Aditya Nigam (SCEE, IIT-Mandi) [email protected] Lecture, February - May, 2019


Siamese CNN - Loss Function

We can use different loss functions for the two types of inputs pairs.
Typical positive pair (xp , xq ) loss: L(xp , xq ) = ||xp − xq ||2
(Euclidean Loss)

Aditya Nigam (SCEE, IIT-Mandi) [email protected] Lecture, February - May, 2019


Siamese CNN - Loss Function

Typical negative pair (xn , xq ) loss:


L(xn , xq ) = max(0, m2 − ||xn − xq ||2 ) (Hinge Loss)

Aditya Nigam (SCEE, IIT-Mandi) [email protected] Lecture, February - May, 2019


Architecture Varieties

Design largely governed by what performs well empirically on the task


at hand.

Aditya Nigam (SCEE, IIT-Mandi) [email protected] Lecture, February - May, 2019

You might also like