Differences

This shows you the differences between two versions of the page.

Link to this comparison view

metric_learning [2017/03/23 22:24]
metric_learning [2017/03/23 22:24] (current)
Line 1: Line 1:
 +====== Metric Learning ======
  
 +
 +Edit: https://​docs.google.com/​a/​codeaudit.com/​document/​d/​1CXcoeD1t34HW6aQHEK8ziHop9UT1RFB-jgRo1B5oTFA/​edit?​usp=sharing
 +
 +Differential Training, Similarity Learning
 +
 +**Discusssion**
 +
 +The goal of metric learning is to ensure that, after training, the distance between vectors of the same class remains small, while distance between different classes are large.
 +
 +**References**
 +
 +http://​citeseerx.ist.psu.edu/​viewdoc/​download?​doi=10.1.1.67.2646&​rep=rep1&​type=pdf ​ DIFFERENTIAL TRAINING OF1
 +ROLLOUT POLICIES
 +
 +http://​www.cs.toronto.edu/​~rsalakhu/​papers/​oneshot1.pdf Siamese Neural Networks for One-shot Image Recognition
 +
 +http://​yann.lecun.com/​exdb/​publis/​pdf/​chopra-05.pdf ​ Learning a Similarity Metric Discriminatively,​ with Application to Face
 +Verification
 +
 +The learning
 +process minimizes a discriminative loss function that drives
 +the similarity metric to be small for pairs of faces from the
 +same person, and large for pairs from different persons.
 +
 +{{https://​ai2-s2-public.s3.amazonaws.com/​figures/​2016-11-01/​6d5e12ee5d75d5f8c04a196dd94173f96dc8603f/​2-Figure1-1.png}}
 +
 +https://​arxiv.org/​abs/​1412.6622 ​ Deep metric learning using Triplet network
 +
 +{{https://​ai2-s2-public.s3.amazonaws.com/​figures/​2016-11-01/​3ac1a7d6630cd1c7b70c4c2c7bc92c40df1162ca/​1-Figure1-1.png}}
 +
 +https://​en.wikipedia.org/​wiki/​Similarity_learning
 +
 +http://​web.cse.ohio-state.edu/​~kulis/​pubs/​ftml_metric_learning.pdf ​ Metric Learning: A Survey
 +
 +http://​arxiv.org/​pdf/​1509.05360v1.pdf Geometry-aware Deep Transform
 +
 +Deep networks are often optimized for a classification
 +objective, where class-labeled samples are input as training
 +; or a metric learning objective, where
 +training data are input as positive and negative pairs.
 +
 +In this section, we first propose a novel deep learning
 +objective that unifies the classification and metric learning
 +criteria. We then introduce a geometry-aware deep transform,
 +and optimize it through standard back-propagation.
 +
 +We denote formulation as Geometry aware
 +Deep Transform (GDT). The GDT objective is a
 +weighted combination of the two formulations. ​  We can understand it as regularizing the
 +metric learning formulation using the classification one.
 +
 +https://​devblogs.nvidia.com/​parallelforall/​understanding-aesthetics-deep-learning/​
 +
 +http://​static.googleusercontent.com/​media/​research.google.com/​en//​pubs/​archive/​41473.pdf ​ DeViSE: A Deep Visual-Semantic Embedding Model
 +
 +One remedy is to leverage data from other
 +sources – such as text data – both to train visual models and to constrain their predictions.
 +In this paper we present a new deep visual-semantic embedding model
 +trained to identify visual objects using both labeled image data as well as semantic
 +information gleaned from unannotated text.
 +
 +{{https://​ai2-s2-public.s3.amazonaws.com/​figures/​2016-11-01/​4aa4069693bee00d1b0759ca3df35e59284e9845/​2-Figure1-1.png}}
 +
 +https://​arxiv.org/​abs/​1610.08904v1 ​ Local Similarity-Aware Deep Feature Embedding
 +
 +the global Euclidean distance cannot faithfully characterize the true feature similarity in a complex visual feature space, where the intraclass distance in a high-density region may be larger than the interclass distance in low-density regions. In this paper, we introduce a Position-Dependent Deep Metric (PDDM) unit, which is capable of learning a similarity metric adaptive to local feature structure.
 +
 +{{https://​ai2-s2-public.s3.amazonaws.com/​figures/​2016-11-01/​30998485c920f62c307c29c4832b70bbce748eaf/​2-Figure2-1.png}}
 +
 +https://​arxiv.org/​pdf/​1611.02268v1.pdf ​ OPTIMAL BINARY AUTOENCODING WITH PAIRWISE
 +CORRELATIONS
 + 
 +https://​arxiv.org/​pdf/​1703.07464v1.pdf No Fuss Distance Metric Learning using Proxies
 +
 +We address the problem of distance metric learning (DML), defined as learning a distance consistent with a notion of semantic similarity. Traditionally,​ for this problem supervision is expressed in the form of sets of points that follow an ordinal relationship -- an anchor point x is similar to a set of positive points Y, and dissimilar to a set of negative points Z, and a loss defined over these distances is minimized. ​
 +While the specifics of the optimization differ, in this work we collectively call this type of supervision Triplets and all methods that follow this pattern Triplet-Based methods. These methods are challenging to optimize. A main issue is the need for finding informative triplets, which is usually achieved by a variety of tricks such as increasing the batch size, hard or semi-hard triplet mining, etc, but even with these tricks, the convergence rate of such methods is slow. In this paper we propose to optimize the triplet loss on a different space of triplets, consisting of an anchor data point and similar and dissimilar proxy points. These proxies approximate the original data points, so that a triplet loss over the proxies is a tight upper bound of the original loss. This proxy-based loss is empirically better behaved. As a result, the proxy-loss improves on state-of-art results for three standard zero-shot learning datasets, by up to 15% points, while converging three times as fast as other triplet-based losses.