What Is Zero-Shot Learning?

Key takeaways

Zero-shot learning (ZSL) trains artificial intelligence (AI) models to identify and classify objects or concepts without prior exposure.

• Zero-shot learning is particularly useful in applications where labeled data is scarce.

• Zero-shot learning is only part of a family of machine learning techniques, which includes variations such as standard zero-shot learning, generalized zero-shot learning, and transductive zero-shot learning. 

• You can gain practical experience with zero-shot learning by working with machine learning frameworks such as PyTorch and TensorFlow.

Learn more about zero-shot learning, its various types, and where it’s impacting the real world. To learn more about the fundamental AI concepts, enroll in the Machine Learning Specialization, where in as little as two months you can develop practical machine learning skills. 

What is zero-shot learning?

Zero-shot learning (ZSL), a machine learning technique, trains AI models to identify and classify objects or concepts without having encountered them before. As a result, the model can achieve its goals without the need for massive amounts of labeled data, which can save time and money. 

In the past, AI models relied on supervised learning, a category of machine learning that trains on labeled data sets and relevant examples to help the model learn. With this method, the models make predictions based on the labeled data they receive in training. The training involves searching the training data and identifying the correct answer, which professionals call the ground truth. The model then learns by teaching itself the difference between the ground truth and its predictions, requiring plenty of labeled data and many hours of training.

In addition to saving time and money, ZSL is useful in fields such as rare disease research and the discovery of unknown species, where the labeled data required for supervised learning may not even exist.

What is zero-shot learning vs. unsupervised learning?

Zero-shot learning enables AI to classify data without direct training examples, whereas unsupervised learning finds hidden patterns in unlabeled data without prior knowledge of class definitions.

Types of zero-shot learning

Zero-shot learning is not a standalone concept; instead, it’s part of a family that includes standard, generalized, and transductive zero-shot learning (among others). None of these ZSL types relies on a particular algorithm or neural network architecture. Instead, they base their existence on addressing the nature of the learning challenge itself by making predictions on unseen, unlabeled data (or “classes”) without having been trained on it before.

It’s worth noting that in ZSL, the “class” in question might have been present in the training data, but it simply appeared unlabeled. This makes large-language models (LLMs) particularly well-suited for ZSL because their training relies on using huge volumes of text for self-supervised learning. It may include incidental (or “auxiliary”) knowledge of unseen classes without it being explicitly labeled as such.

Learn more about the different types of zero-shot learning, its uses, and some potential advantages and drawbacks.

Read more: What Is Neural Network Classification? Techniques and Applications

1. Standard zero-shot learning

Standard zero-shot learning relies on transferring knowledge from familiar to unfamiliar classes. It involves using known semantic or logical information, such as text, labels, or descriptions, to create a new space for unseen classes.

In standard ZSL, the model maps out a semantic space to better understand the relationships between seen and unseen data. From there, it can make logical connections and apply the same thinking to classify new data into unseen categories. This trains the model to predict and classify data from unseen classes more accurately going forward.

2. Generalized zero-shot learning

Generalized zero-shot learning (GZSL) takes training a step further than standard ZSL by incorporating both seen and unseen classes into the mix. Unlike standard ZSL, which focuses only on unseen classes, GZSL is more true to life by presenting the model with new information alongside data it might have encountered before.

The goal is to address a common bias in standard ZSL models: the model tends to favor seen classes over unseen ones. GSZL works to reduce the bias toward the seen and improve performance for the unseen.

3. Transductive zero-shot learning

Transductive zero-shot learning goes one step further than GZSL by utilizing unseen classes and unlabeled data points during the training phase. GZSL attempts to give the model more lifelike training scenarios by incorporating seen and unseen classes, but it still provides the model with data that logically matches.

In transductive zero-shot learning scenarios, the data doesn’t match as logically as with generalized or standard ZSL training sessions. This additional challenge helps improve the model’s ability to generalize, thus allowing it to make better predictions when faced with real-world tasks.

What does zero-shot learning do?

Zero-shot learning aims to train artificial intelligence models to recognize new objects by identifying similarities between seen and unseen classes. This helps the model make more informed guesses, and, similar to human intuition, this methodology helps the model adapt and learn new information. 

Instead of relying solely on seen classes, zero-shot learning teaches the model how to make assumptions based on key defining features. For example, the model might know that cats have fur, whiskers, and a tail, allowing it to infer that an image of a dog with a wagging tail and a wet snout is likely unrelated to a feline.

This is handy in real-world applications where examples might not exist (smaller amounts of labeled, organized data). Think of the potential for ZSL in natural language processing, where models could translate between languages thanks to their knowledge of the relationships between dialects. In the medical field, health care providers could use zero-shot learning to detect new conditions based on the model’s understanding of similar concepts. 

Who uses zero-shot learning?

As you might imagine, various industries use zero-shot learning, including academia and research, the technology and AI industry, and media and content management.

Academia and research

Learners at the University of Nebraska, Lincoln, and at Lehigh University are exploring zero-shot learning techniques as part of the academic institution’s research on learner performance in science, technology, engineering, and mathematics (STEM) education. Using ZSL, they hope to find a way to predict end-of-semester grades based on learners’ early work in STEM courses. It’s a true test of ZSL’s ability to make accurate forecasts based on a few unseen classes.

Technology and the AI industry

OpenAI, the artificial intelligence technology company behind ChatGPT, used zero-shot learning to power its neural network, CLIP. The model trained on a wide range of images and natural language processing information from the internet, which allowed CLIP to learn the differences between pictures and text descriptions through zero-shot learning.

Media and content management

Streaming service Netflix has begun researching the use of zero-shot learning to make media and content recommendations to users. In their preliminary study, researchers found that zero-shot large-language models made better recommendations than non-ZSL models.

Pros and cons of using zero-shot learning

Zero-shot learning has both advantages and challenges. Learn ZSL’s most prominent strengths and weaknesses.

Advantages

• Increased efficiency: ZSL models can automatically handle new, unseen classes without retraining, enabling faster deployment and easier adaptation with less downtime.

• Improved generalization: ZSL models are highly effective at generalizing from seen to unseen classes, allowing the LLMs to make inferences without direct examples.

Challenges

• Reduced performance: Because ZSL models make inferences based on unseen classes, a risk of incorrect generalization exists, especially when the unseen classes are drastically different from the training data. This can lead to performance issues that waste time instead of saving it.

• Greater complexity: Due to the vast, varied, and occasionally illogical nature of real-world data, ZSL models might excel during standard zero-shot learning scenarios but struggle in generalized or transductive settings. 

How to get started in zero-shot learning

Getting started in zero-shot learning begins with a foundational understanding of artificial intelligence and computer science, specifically, the basics of machine learning, neural networks, and natural language processing. Beyond this, you’ll also want to gain practical experience with PyTorch, TensorFlow, and other machine learning frameworks. It also helps to complete courses in artificial intelligence ethics and fairness because ZSL models can reflect harmful biases pulled from the training data.

Your go-to machine learning guide

Discover fresh insights into your field or learn about trends in your industry by subscribing to our LinkedIn newsletter, Career Chat! Or if you want to keep learning more about AI and machine learning, check out these free resources:

• Watch on YouTube: Machine Learning in Real Life: From Spotify to Healthcare

• Discover career paths: Machine Learning Career Paths: Explore Roles and Specializations

• Explore learning options: Machine Learning Roadmap: Beginner to Expert

Accelerate your career growth with a Coursera Plus subscription. When you enroll in either the monthly or annual option, you’ll get access to over 10,000 courses.

Build job-ready skills with Coursera Plus

Start 7-day free trial

• 
• 
• 
• 

Start 7-day free trial