Yuval Atzmon 1,2, Gal Chechik 1,2
1Bar-Ilan University, Israel 2NVIDIA Research
Adaptive Confidence Smoothing for Generalized Zero-Shot Learning
In ZSL we learn new (unseen) classes from a description, without any visual examples
Zero Shot Learning (ZSL)
Could you recognize a Jackalope?
“A Jackalope is a rabbit with horns.”
Generalized Zero Shot Learning (GZSL)
Idea 1: Soft combination of expertsBreak the model to domain experts. Inspired by the dual-route reading model in cognitive psychology.
(1) Seen-classes expert(2) Unseen-classes expert (a ZSL model)(3) Gater: S/U classifier ,
=class, =Seen, =Unseen
The full modeladaptive COnfidence SMOothing (COSMO)
Step 1: Experts Step 2: GaterStep 3: Smoothing Step 4: Combine
Our approach (COSMO, blue) outperforms previous non-GAN approaches (triangles) and generative approaches (crosses).
Acc
ura
cy U
nse
en [
%]
Accuracy Seen [%]
Seen - Unseen accuracy curveby sweeping gate decision boundary
Image In GZSL, at test time, we can either see an image from a seen class or from a new unseen class. Gater is “aware” of experts response
“I’m cooking tonight and you can rely on me to absquatulate the moment it’s done.”“I’m cooking tonight and you can rely on me to absquatulate the moment it’s done.”
Seen expert
Paper, code and video:http://bit.ly/COSMO123
In reading, once we encounter an unfamiliar word, we
compose it from syllables.
IdealGenerative
COSMO+LAGO(ours)
LAGO (baseline)
CS+LAGO
COSMO
CUB: Fine grained bird recognition
AWA: Animal recognition
SUN: Visual scenes
~12K images 150/50 S/U classes
~37K images 40/10 S/U classes
~14K images 645/72 S/U classes
Train: Seen classes Test: Unseen
Rabbit: rodent-shape with long ears
Puku: antelope with ridge-structured horns
or Seen
Jackalope: rabbit with horns
Saiga: antelope with bloated nostrils
Rabbit: ...
Puku: ...
Standard ZSL models fail in GZSL due to
• Spurious correlations
• Domain adaptation
• Extremely imbalanced data
Three benchmark datasetsOur solution: Use a uniform prior during inference, with adaptive weight λ=p(S)=1-p(U), set by the gater belief
Idea 2: Smooth over-confident experts
Over-confidentprediction
uniform prior
Images outside-the-domain of an expert, usually produce over-confident predictions.
Instead, all classes should have uniformly low probabilities, since they are all ”equally wrong”.
GZSL requires robustness across Seen/Unseen domains.
COSMO softly combines domain experts and smooths their predictions to address over-confident experts.
With COSMO, standard ZSL classifiers can outperform generative classifiers.
Takeaways
LAGO zero-shot expert(Atzmon, 2018)
Participation in CVPR is supported by the Israeli ministry of science
a seen class or from a new unseen class. Gater is “aware” of experts response
The gater is trained to discriminate the response of experts to seen and unseen images.
Top-K pooling to achieve invariance to order and identity of input classes.
Learns new attribute compositions, with a differentiable AND-OR architecture
Desired smooth prediction
Gater: S/UClassifier
Unseen (ZSL) expert
Confidence Based Gating