Source-linked AI summary
A Simple Exponential Family Framework for Zero-Shot Learning
Vinay Kumar Verma, Piyush Rai
TL;DR
Zero-shot learning must recognize classes without labeled examples from those classes, while existing approaches may not adequately represent intra-class variability. The paper introduces an exponential-family generative framework whose distribution parameters are predicted from class attributes using regression learned on seen classes. Across benchmark datasets, the framework improves inductive ZSL over baselines and extends to unlabeled-data and few-shot settings, while generalized few-shot learning remains an open direction.
Problem
Zero-shot learning requires recognizing unseen classes without labeled training examples, and fixed-point semantic embeddings do not adequately account for intra-class variability.
Method
The framework models each seen or unseen class with an exponential-family class-conditional distribution whose parameters are predicted from class attributes by learned linear or nonlinear regression.
Results
On CUB-200 and SUN, both linear and nonlinear models outperform reported state-of-the-art baselines; across three datasets, the average improvement over the overall best baseline is 5.67%.
Takeaways & Limitations
The framework supports transductive zero-shot learning and few-shot learning, including unlabeled unseen-class data and small numbers of labeled unseen-class examples.
Takeaways & Limitations
The authors leave more robust handling of generalized few-shot learning as future work and identify fully Bayesian parameter learning as another open direction.
Abstract
from arXiv · showhide
We present a simple generative framework for learning to predict previously unseen classes, based on estimating class-attribute-gated class-conditional distributions. We model each class-conditional distribution as an exponential family distribution and the parameters of the distribution of each seen/unseen class are defined as functions of the respective observed class attributes. These functions can be learned using only the seen class data and can be used to predict the parameters of the class-conditional distribution of each unseen class. Unlike most existing methods for zero-shot learning that represent classes as fixed embeddings in some vector space, our generative model naturally represents each class as a probability distribution. It is simple to implement and also allows leveraging additional unlabeled data from unseen classes to improve the estimates of their class-conditional distributions using transductive/semi-supervised learning. Moreover, it extends seamlessly to few-shot learning by easily updating these distributions when provided with a small number of additional labelled examples from unseen classes. Through a comprehensive set of experiments on several benchmark data sets, we demonstrate the efficacy of our framework.
1 Introduction
Zero-shot learning targets recognition of classes without labeled training examples, motivating robust approaches that handle scarce supervision. The paper proposes a generative framework that models class-conditional distributions through class attributes and supports unlabeled and few-shot extensions.
- Motivation: Zero-shot learning recognizes objects from classes absent during training, especially when new concepts must be learned without labeled novel-class data.Such settings motivate methods that remain robust and scalable under scarce supervision.
- Limitations of Existing Methods: Embedding-based ZSL methods represent each class as a fixed semantic-space point, which does not adequately capture intra-class variability.
- Motivation: A generative model can represent different data types with class-conditional distributions and incorporate unlabeled data during parameter estimation.The paper identifies transductive or semi-supervised estimation as useful when labeled data are scarce or seen and unseen distributions differ.
- Proposed Framework: The proposed framework models seen and unseen classes with exponential-family distributions whose parameters are conditioned on class-attribute vectors through linear or nonlinear regression.The learned regression predicts unseen-class distribution parameters using only their attributes.
- Practical Extensions: Parameter estimation reduces to a regression problem with a closed-form solution, while unlabeled unseen data and a few labeled unseen examples can be incorporated through conjugate-prior distributions.
2 A Generative Framework For ZSL
The framework models seen and unseen class-conditional distributions with exponential-family parameters gated by class attributes, then learns attribute-to-parameter regressions from seen classes to estimate unseen classes. Its generative, modular design also supports nonlinear models, unlabeled unseen-class data, other exponential families, and online few-shot updates.
- Each class-conditional distribution is modeled as an exponential family with natural parameters, sufficient statistics, and a log-partition function.
- ZSL provides only labelled training examples for seen classes, so unseen-class conditional-distribution parameters cannot be estimated directly.
- 2.1 Gating via Class-Attributes: Class attributes are mapped to distribution parameters through regression functions, allowing unseen-class parameters to be predicted from their attributes.The framework learns separate functions for Gaussian means and variances, using linear or nonlinear regression choices.
- 2.2 Learning The Regression Functions: For Gaussian class-conditionals, the method estimates seen-class parameters from data and fits regularized linear or kernel regressions that produce unseen-class parameters.The nonlinear model uses kernel similarities between unseen and seen class attributes without explicitly computing the feature mapping.
- The same framework generalizes beyond Gaussian distributions because its parameter-estimation recipe applies to other exponential-family distributions.
- Generative parameter estimates can incorporate unlabeled unseen-class data and a small number of labelled unseen examples, with few-shot updates available online.Unlabeled data supports transductive or semi-supervised estimation, while additional labelled examples can update the model without retraining from scratch.
3 Related Work
Related ZSL methods use semantic embeddings, combinations of classes, or unlabeled data, while this framework represents each class with a distribution and supports general unlabeled-data integration.
- Embedding methods represent seen and unseen classes as points in a shared semantic space, limiting their ability to model intra-class variability.
- This framework instead models each class with its own distribution, providing greater flexibility for capturing intra-class variability.
- Some ZSL approaches model unseen classes as linear or convex combinations of seen or basis classes; the linear version here subsumes basis-class models as a special case.
- Domain shift can occur when seen and unseen classes come from different domains, and unlabeled unseen-class data can help alleviate it.
- Unlike model-specific semi-supervised approaches, this framework provides a general and simple mechanism for incorporating unlabeled data, with efficient solvers available for large-scale problems.
4 Experiments
The experiments evaluate GFZSL across standard, transductive, few-shot, and generalized few-shot settings on benchmark datasets using established image features and class-attribute splits. Results show strong inductive and transductive performance, substantial few-shot gains, and a remaining challenge in generalized few-shot learning.
- Experimental setup: GFZSL is evaluated on AwA, CUB-200, and SUN using standard dataset splits, class attributes, and VGG-19 or GoogleNet image features.AwA has 40 seen and 10 unseen classes; CUB-200 has 150 seen and 50 unseen classes; SUN has 707 seen and 10 unseen classes.
- Experimental setup: The experiments cover inductive ZSL using seen-class labels, transductive ZSL using unseen-class unlabeled data, few-shot learning using a small number of unseen-class labels, and generalized ZSL.Generalized ZSL allows test examples from both seen and unseen classes.
- Inductive ZSL: On CUB-200 and SUN, both linear and nonlinear GFZSL models outperform all compared state-of-the-art methods in the inductive setting.On AwA, the model is marginally below the best baseline, while averaging a 5.67% improvement across all three datasets over the overall best baseline.
- Inductive ZSL: 21.05% relative improvement is achieved over the best VGG-19 baseline on CUB-200 in the inductive setting.CUB-200 is described as difficult because it contains many fine-grained classes.
- Transductive ZSL: The generative classifier evaluates each test example under class-conditional distributions, accounting for their shape and spread rather than using point-based semantic-space similarity.In the transductive setting, unseen-class distributions are estimated using unlabeled data before assigning each example to the highest-probability class.
- Few-shot learning: Few-shot accuracy significantly exceeds standard inductive ZSL with only 2 or 5 additional labeled examples per unseen class, and the method outperforms multiclass SVM.The improvement is reported across the evaluated datasets and is attributed to knowledge transfer from seen-class data.
- Generalized few-shot learning: In generalized few-shot learning, accuracy improves as unseen-class labels are added, but the model remains below multiclass SVM.The authors identify the need for more robust handling of this setting as future work.
5 Conclusion
The framework models classes as attribute-conditioned exponential-family distributions and is designed for simple, modular extension to transductive and few-shot settings. The authors identify jointly learning class attributes and fully Bayesian parameter learning as future directions.
- The framework represents each seen and unseen class as a distribution whose parameters depend on its class-attribute vector.
- Its simple, modular architecture supports interchangeable algorithms and natural extensions to transductive zero-shot and few-shot learning.
- The framework is easy to implement and can scale to many classes using advances in large-scale regression.
- Future work could jointly learn class attributes from external data or adopt fully Bayesian learning instead of point parameter estimation.