Roelof Pieters
“Zero-­‐Shot	
  Learning	
  Through	
  
Cross-­‐Modal	
  Transfer”
20	
  February	
  2015	
  

Deep	
  Learning	
  Reading	
  Group	
  
Richard Socher, Milind Ganjoo, Hamsa Sridhar, Osbert Bastani,
Christopher D. Manning, Andrew Y. Ng
http://arxiv.org/abs/1301.3666
@graphific
ICLR 2013
http://papers.nips.cc/… NIPS 2013
Review of
http://www.csc.kth.se/~roelof/
“a zero-shot model that can predict
both seen and unseen classes”
Zero-Shot Learning Through Cross-Modal Transfer

Richard Socher, Milind Ganjoo, Hamsa Sridhar, Osbert
Bastani, Christopher D. Manning, Andrew Y. Ng, ICLR 2013
Core (novel) Idea:
Key Ideas
• Semantic word vector representations:
• Allows transfer of knowledge between modalities
• Even when these representations are learned in an
unsupervised way
• Bayesian framework:
1. Differentiate between unseen/seen classes
2. From points on the semantic manifold of trained classes
3. Allows combining both zero-shot and seen classification
into one framework

[ Ø-shot + 1-shot = “multi-shot” :) ]
Visual-Semantic Word Space
• Word vectors capture distributional similarities
from a large, unsupervised text corpus. [Word
vectors create a semantic space] (Huang et al. 2012)
• Images are mapped into a semantic space of words that is
learned by a neural network model (Coates & Ng 2011;
Coates et al. 2011)
• By learning an image mapping into this space, the word
vectors get implicitly grounded by the visual modality,
allowing us to give prototypical instances for various word
(Socher et al. 2013, this paper)
• Word vectors capture distributional similarities from a
large, unsupervised text corpus. [Word vectors create a
semantic space] (Huang et al. 2012)
• Images are mapped into a semantic space of words that is
learned by a neural network model (Coates & Ng 2011;
Coates et al. 2011)
• By learning an image mapping into this space, the word
vectors get implicitly grounded by the visual modality,
allowing us to give prototypical instances for various word
(Socher et al. 2013, this paper)
Visual-Semantic Word Space
Semantic Space
Visual-Semantic Word Space
Semantic Space
E. H. Huang, R. Socher, C. D. Manning, andA. Y. Ng.
Improving Word Representations via Global Context and
Multiple Word Prototypes. InACL, 2012
• (Socher et al. 2013) uses pre-trained (50-d) word
vectors of (Huang et al. 2012):
http://www.socher.org/index.php/Main/ImprovingWordRepresentationsViaGlobalContextAndMultipleWordPrototypes
[Huang et al. 2012]
Semantic Space
“You shall know a word by the company it keeps”

(J. R. Firth 1957)
One of the most successful ideas of modern
statistical NLP!
these words represent banking
Distributed Semantics (short recap)
[Huang et al. 2012]
[Huang et al. 2012]
local context:
global context:
final score:
[Huang et al. 2012]
activation of the hidden layer with h hidden nodes
1st layer weights
2th layer weights
1st layer bias
2th layer bias
local context
scoring function
element-wise activation function (ie tanh)
concatenation of the m word embeddings
representing sequence s
activation of the hidden layer with h(g) hidden nodes
1st layer weights
2th layer weights
1st layer bias
2th layer bias
concatenation of the weighted average document
vector and the vector of the last word in s
weighting function that captures the importance of
word ti in the document (tf-idf)
global context
scoring function
document(s) as ordered list of word embeddings
weighted average of all word vectors in a document
• Word vectors capture distributional similarities from a
large, unsupervised text corpus. [Word vectors create a
semantic space] (Huang et al. 2012)
• Images are mapped into a semantic space of words
that is learned by a neural network model (Coates &
Ng 2011; Coates et al. 2011)
• By learning an image mapping into this space, the word
vectors get implicitly grounded by the visual modality,
allowing us to give prototypical instances for various word
(Socher et al. 2013, this paper)
Visual-Semantic Word Space
Semantic Space
• Word vectors capture distributional similarities from a
large, unsupervised text corpus. [Word vectors create a
semantic space] (Huang et al. 2012)
• Images are mapped into a semantic space of words that is
learned by a neural network model (Coates & Ng 2011;
Coates et al. 2011)
• By learning an image mapping into this space, the word
vectors get implicitly grounded by the visual modality,
allowing us to give prototypical instances for various word
(Socher et al. 2013, this paper)
Visual-Semantic Word Space
Semantic Space
Image Features
Image Feature Learning
• high level description: extract random patches, extract features
from sub-patches, pool features, train liner classifier to predict
labels
• = fast simple algorithms with the correct parameters work as
well as complex, slow algorithms
[Coates et al. 2011 (used by Coates & Ng 2011)]
• Word vectors capture distributional similarities from a
large, unsupervised text corpus. [Word vectors create a
semantic space] (Huang et al. 2012)
• Images are mapped into a semantic space of words that is
learned by a neural network model (Coates & Ng 2011;
Coates et al. 2011)
• By learning an image mapping into this space, the
word vectors get implicitly grounded by the visual
modality, allowing us to give prototypical instances
for various word (Socher et al. 2013, this paper)
Visual-Semantic Word Space
Semantic Space
Image Features
• Word vectors capture distributional similarities from a
large, unsupervised text corpus. [Word vectors create a
semantic space] (Huang et al. 2012)
• Images are mapped into a semantic space of words that is
learned by a neural network model (Coates & Ng 2011;
Coates et al. 2011)
• By learning an image mapping into this space, the word
vectors get implicitly grounded by the visual modality,
allowing us to give prototypical instances for various word
(Socher et al. 2013, this paper)
Visual-Semantic Word Space
Semantic Space
Visual-Semantic Space
Image Features
[this paper, Socher et al. 2013]
Visual-Semantic Space
Projecting Images into Visual Space
Objective function(s):
[Socher et al. 2013]
training images
set of word vectors seen/unseen visual classes
mapped to the word vector (class name)
Projecting Images into Visual Space
Objective function(s):
[Socher et al. 2013]
training images
set of word vectors seen/unseen visual classes
mapped to the word vector (class name)
T-SNE visualization of the semantic word space [Socher et al. 2013]
[Socher et al. 2013]
Projecting Images into Visual Space
Mapped points of
seen classes:
(Outlier Detection)
Predicting class y:
binary visibility random variable
probability of an image being in an unseen class
Treshold T:
[Socher et al. 2013]
Projecting Images into Visual Space
(Outlier Detection)
binary visibility random variable
probability of an image being in an unseen class
known class prediction:
[Socher et al. 2013][Socher et al. NIPS 2013]
Results
Main Contributions
Zero-shot learning
• Good classification of (pairs of) unseen classes can be
achieved based on learned representations for these
classes
• => as opposed to hand designed representations
• => extends (Lampert 2009; Guo-Jun 2011) [Manual defined
visual/semantic attributes to classify unseen classes]
Main Contributions
“Multi”-shot learning
• Deal with both seen and unseen classes: Allows combining both
zero-shot and seen classification into one framework:

[ Ø-shot + 1-shot = “multi-shot” :) ]
• Assumption: unseen classes as outliers
• Major weakness:

drop from 80% to 70% for 15%-30% accuracy (on particular classes)
• => extends (Lampert 2009; Palatucci 2009) [manual defined
representations, limited to zero-shot classes], using outlier
detection
• => extends (Weston et al. 2010) (joint embedding images and labels
through linear mapping) [linear mapping only, so cant generalise to
new classes: 1-shot], using outlier detection
Main Contributions
Knowledge-Transfer
• Allows transfer of knowledge between modalities, within
multimodal embeddings
• Allows for unsupervised matching
• => extends (Socher & Fei-Fei 2012) (kernelized canonical
correlation analysis) [still require small amount of training
data for each class: 1-shot]
• => extends (Salakhutdinov et al. 2012) (learn low-level image
features followed by a probabilistic model to transfer
knowledge) [also limited to 1-shot classes]
Bibliography
• C. H. Lampert, H. Nickisch, and S. Harmeling. Learning to
Detect Unseen Object Classes by Between-Class Attribute
Transfer. In CVPR, 2009
• M. Palatucci, D. Pomerleau, G. Hinton, and T. Mitchell. Zero-
shot learning with semantic output codes. In NIPS, 2009
• Guo-Jun Qi, C. Aggarwal, Y. Rui, Q. Tian, S. Chang, and T. Huang.
Towards cross-category knowledge propagation for learning visual
concepts. In CVPR, 2011
• R. Socher and L. Fei-Fei. Connecting modalities: Semi-supervised
segmentation and annotation of images using unaligned text
corpora. In CVPR, 2010
• E. H. Huang, R. Socher, C. D. Manning, and A. Y. Ng. Improving
Word Representations via Global Context and Multiple Word
Prototypes. In ACL, 2012
Bibliography
• A. Coates and A. Ng. The Importance of Encoding Versus
Training with Sparse Coding and Vector Quantization. In ICML,
2011.
• Coates, Adam, Lee, Honlak, and Ng, Andrew Y. An analysis of
single-layer networks in unsupervised feature learning. In
International Conference on AI and Statistics, 2011.
• A. Torralba R. Salakhutdinov, J. Tenenbaum. Learning to learn
with compound hierarchical-deep models. In NIPS, 2012.
• J. Weston, S. Bengio1, and N. Usunier. Large Scale Image
Annotation: Learning to Rank with Joint Word-Image
Embeddings. In Machine Learning, 81 (1):21-35, 2010

Zero shot learning through cross-modal transfer

  • 1.
    Roelof Pieters “Zero-­‐Shot  Learning  Through   Cross-­‐Modal  Transfer” 20  February  2015  
 Deep  Learning  Reading  Group   Richard Socher, Milind Ganjoo, Hamsa Sridhar, Osbert Bastani, Christopher D. Manning, Andrew Y. Ng http://arxiv.org/abs/1301.3666 @graphific ICLR 2013 http://papers.nips.cc/… NIPS 2013 Review of http://www.csc.kth.se/~roelof/
  • 2.
    “a zero-shot modelthat can predict both seen and unseen classes” Zero-Shot Learning Through Cross-Modal Transfer
 Richard Socher, Milind Ganjoo, Hamsa Sridhar, Osbert Bastani, Christopher D. Manning, Andrew Y. Ng, ICLR 2013 Core (novel) Idea:
  • 3.
    Key Ideas • Semanticword vector representations: • Allows transfer of knowledge between modalities • Even when these representations are learned in an unsupervised way • Bayesian framework: 1. Differentiate between unseen/seen classes 2. From points on the semantic manifold of trained classes 3. Allows combining both zero-shot and seen classification into one framework
 [ Ø-shot + 1-shot = “multi-shot” :) ]
  • 4.
    Visual-Semantic Word Space •Word vectors capture distributional similarities from a large, unsupervised text corpus. [Word vectors create a semantic space] (Huang et al. 2012) • Images are mapped into a semantic space of words that is learned by a neural network model (Coates & Ng 2011; Coates et al. 2011) • By learning an image mapping into this space, the word vectors get implicitly grounded by the visual modality, allowing us to give prototypical instances for various word (Socher et al. 2013, this paper)
  • 5.
    • Word vectorscapture distributional similarities from a large, unsupervised text corpus. [Word vectors create a semantic space] (Huang et al. 2012) • Images are mapped into a semantic space of words that is learned by a neural network model (Coates & Ng 2011; Coates et al. 2011) • By learning an image mapping into this space, the word vectors get implicitly grounded by the visual modality, allowing us to give prototypical instances for various word (Socher et al. 2013, this paper) Visual-Semantic Word Space Semantic Space
  • 6.
    Visual-Semantic Word Space SemanticSpace E. H. Huang, R. Socher, C. D. Manning, andA. Y. Ng. Improving Word Representations via Global Context and Multiple Word Prototypes. InACL, 2012 • (Socher et al. 2013) uses pre-trained (50-d) word vectors of (Huang et al. 2012):
  • 7.
  • 8.
    “You shall knowa word by the company it keeps”
 (J. R. Firth 1957) One of the most successful ideas of modern statistical NLP! these words represent banking Distributed Semantics (short recap)
  • 9.
  • 10.
  • 11.
    local context: global context: finalscore: [Huang et al. 2012]
  • 12.
    activation of thehidden layer with h hidden nodes 1st layer weights 2th layer weights 1st layer bias 2th layer bias local context scoring function element-wise activation function (ie tanh) concatenation of the m word embeddings representing sequence s
  • 13.
    activation of thehidden layer with h(g) hidden nodes 1st layer weights 2th layer weights 1st layer bias 2th layer bias concatenation of the weighted average document vector and the vector of the last word in s weighting function that captures the importance of word ti in the document (tf-idf) global context scoring function document(s) as ordered list of word embeddings weighted average of all word vectors in a document
  • 14.
    • Word vectorscapture distributional similarities from a large, unsupervised text corpus. [Word vectors create a semantic space] (Huang et al. 2012) • Images are mapped into a semantic space of words that is learned by a neural network model (Coates & Ng 2011; Coates et al. 2011) • By learning an image mapping into this space, the word vectors get implicitly grounded by the visual modality, allowing us to give prototypical instances for various word (Socher et al. 2013, this paper) Visual-Semantic Word Space Semantic Space
  • 15.
    • Word vectorscapture distributional similarities from a large, unsupervised text corpus. [Word vectors create a semantic space] (Huang et al. 2012) • Images are mapped into a semantic space of words that is learned by a neural network model (Coates & Ng 2011; Coates et al. 2011) • By learning an image mapping into this space, the word vectors get implicitly grounded by the visual modality, allowing us to give prototypical instances for various word (Socher et al. 2013, this paper) Visual-Semantic Word Space Semantic Space Image Features
  • 16.
    Image Feature Learning •high level description: extract random patches, extract features from sub-patches, pool features, train liner classifier to predict labels • = fast simple algorithms with the correct parameters work as well as complex, slow algorithms [Coates et al. 2011 (used by Coates & Ng 2011)]
  • 17.
    • Word vectorscapture distributional similarities from a large, unsupervised text corpus. [Word vectors create a semantic space] (Huang et al. 2012) • Images are mapped into a semantic space of words that is learned by a neural network model (Coates & Ng 2011; Coates et al. 2011) • By learning an image mapping into this space, the word vectors get implicitly grounded by the visual modality, allowing us to give prototypical instances for various word (Socher et al. 2013, this paper) Visual-Semantic Word Space Semantic Space Image Features
  • 18.
    • Word vectorscapture distributional similarities from a large, unsupervised text corpus. [Word vectors create a semantic space] (Huang et al. 2012) • Images are mapped into a semantic space of words that is learned by a neural network model (Coates & Ng 2011; Coates et al. 2011) • By learning an image mapping into this space, the word vectors get implicitly grounded by the visual modality, allowing us to give prototypical instances for various word (Socher et al. 2013, this paper) Visual-Semantic Word Space Semantic Space Visual-Semantic Space Image Features
  • 19.
    [this paper, Socheret al. 2013] Visual-Semantic Space
  • 20.
    Projecting Images intoVisual Space Objective function(s): [Socher et al. 2013] training images set of word vectors seen/unseen visual classes mapped to the word vector (class name)
  • 21.
    Projecting Images intoVisual Space Objective function(s): [Socher et al. 2013] training images set of word vectors seen/unseen visual classes mapped to the word vector (class name)
  • 22.
    T-SNE visualization ofthe semantic word space [Socher et al. 2013]
  • 23.
    [Socher et al.2013] Projecting Images into Visual Space Mapped points of seen classes: (Outlier Detection) Predicting class y: binary visibility random variable probability of an image being in an unseen class Treshold T:
  • 24.
    [Socher et al.2013] Projecting Images into Visual Space (Outlier Detection) binary visibility random variable probability of an image being in an unseen class known class prediction:
  • 25.
    [Socher et al.2013][Socher et al. NIPS 2013] Results
  • 26.
    Main Contributions Zero-shot learning •Good classification of (pairs of) unseen classes can be achieved based on learned representations for these classes • => as opposed to hand designed representations • => extends (Lampert 2009; Guo-Jun 2011) [Manual defined visual/semantic attributes to classify unseen classes]
  • 27.
    Main Contributions “Multi”-shot learning •Deal with both seen and unseen classes: Allows combining both zero-shot and seen classification into one framework:
 [ Ø-shot + 1-shot = “multi-shot” :) ] • Assumption: unseen classes as outliers • Major weakness:
 drop from 80% to 70% for 15%-30% accuracy (on particular classes) • => extends (Lampert 2009; Palatucci 2009) [manual defined representations, limited to zero-shot classes], using outlier detection • => extends (Weston et al. 2010) (joint embedding images and labels through linear mapping) [linear mapping only, so cant generalise to new classes: 1-shot], using outlier detection
  • 28.
    Main Contributions Knowledge-Transfer • Allowstransfer of knowledge between modalities, within multimodal embeddings • Allows for unsupervised matching • => extends (Socher & Fei-Fei 2012) (kernelized canonical correlation analysis) [still require small amount of training data for each class: 1-shot] • => extends (Salakhutdinov et al. 2012) (learn low-level image features followed by a probabilistic model to transfer knowledge) [also limited to 1-shot classes]
  • 29.
    Bibliography • C. H.Lampert, H. Nickisch, and S. Harmeling. Learning to Detect Unseen Object Classes by Between-Class Attribute Transfer. In CVPR, 2009 • M. Palatucci, D. Pomerleau, G. Hinton, and T. Mitchell. Zero- shot learning with semantic output codes. In NIPS, 2009 • Guo-Jun Qi, C. Aggarwal, Y. Rui, Q. Tian, S. Chang, and T. Huang. Towards cross-category knowledge propagation for learning visual concepts. In CVPR, 2011 • R. Socher and L. Fei-Fei. Connecting modalities: Semi-supervised segmentation and annotation of images using unaligned text corpora. In CVPR, 2010 • E. H. Huang, R. Socher, C. D. Manning, and A. Y. Ng. Improving Word Representations via Global Context and Multiple Word Prototypes. In ACL, 2012
  • 30.
    Bibliography • A. Coatesand A. Ng. The Importance of Encoding Versus Training with Sparse Coding and Vector Quantization. In ICML, 2011. • Coates, Adam, Lee, Honlak, and Ng, Andrew Y. An analysis of single-layer networks in unsupervised feature learning. In International Conference on AI and Statistics, 2011. • A. Torralba R. Salakhutdinov, J. Tenenbaum. Learning to learn with compound hierarchical-deep models. In NIPS, 2012. • J. Weston, S. Bengio1, and N. Usunier. Large Scale Image Annotation: Learning to Rank with Joint Word-Image Embeddings. In Machine Learning, 81 (1):21-35, 2010