Research Focus

Zero-Shot Learning

Learning without examples through semantic knowledge transfer

What is Zero-Shot Learning?

Zero-Shot Learning (ZSL) is a machine learning paradigm that enables models to classify objects that they have never seen during training. To do this, it is based on the transfer of semantic knowledge, using high-level descriptions (attributes, text embeddings, or knowledge graphs) to relate known classes to unknown classes. semantic relationships that link known classes with unknown ones.

The set of all those descriptions makes up the key component, the semantic space.

Types of semantic space?

According to its nature, semantic spaces can be classified as:

Implicit

It is the semantic space that comes implicit in large language models. It is a space originated from the vast amount of data with which these models have been trained. It is a space very rich in knowledge, but difficult to interpret.

Explicit

Manual

It is a semantic space defined manually by human experts. It consists of attributes that describe images, e.g., "has wings", "is red", etc. This type of space is interpretable but costly to generate.

Explicit

Learned

Semantic spaces generated by the output of a machine learning model. They are more flexible and easier to generate, but may lack interpretability and precision.

Types of Methods

There are two main approaches to addressing ZSL tasks:

Generative Methods

Synthetic samples of unseen classes are generated by conditioning a generative model on the semantic space of those classes. Subsequently, a supervised classifier is trained with both real and synthetic samples.

Attention-based Methods

Through attention mechanisms, the model learns to identify attributes from the semantic space in the image. This information is then used to classify instances of unseen classes.

Current Challenges

Despite its progress, Zero-Shot Learning faces several fundamental challenges that define current research lines

Domain Shift

Discrepancy between the distribution of seen and unseen classes, causing the model to tend to predict seen classes even for instances of unseen classes.

Hubness Problem

In high-dimensional spaces, certain points ("hubs") become frequent neighbors of many others, generating biased classifications.

Semantic Space Quality

The effectiveness of ZSL critically depends on the quality of semantic descriptions, which can be noisy, incomplete, or subjective.

Bias in GZSL

Models tend to strongly favor seen classes over unseen ones, making balanced performance in generalized scenarios difficult.

Our Research

In the GPAIS group at IAFER Chair, we are working on innovative Zero-Shot Learning techniques that address these fundamental challenges:

🚀 Active Research Lines

Semantic Space Preprocessing: Techniques to clean and improve the quality of the semantic space used in ZSL
ZSL using Fuzzy Logic: Using fuzzy logic to handle uncertainty and ambiguity in zero-shot classification

📚 Publications

Loading publications...

Practical Applications

Zero-Shot Learning has a significant impact in multiple domains where it is impossible or impractical to obtain training examples for all possible categories:

Rare species recognition: Classification of endangered animals without abundant photographs
Medical diagnosis: Identification of rare diseases with few documented cases
Object recognition in e-commerce: Classification of millions of products without manually labeling each category
Activity detection: Recognition of complex or unusual human actions in video surveillance
Semantic image search: Finding images based on textual descriptions without prior examples
Robotics and object manipulation: Robots that recognize and manipulate new objects without retraining