A new breed of artificial intelligence is emerging, capable of performing tasks that were never explicitly taught. Known as zero-shot learning, this technology is pushing the boundaries of AI capabilities and raising questions about the future of machine intelligence.
For example, OpenAI’s GPT-4 is a language model that demonstrates zero-shot learning skills. Without any specific legal training, GPT-4 scored in the 90th percentile on the Uniform Bar Exam. The model also had zero-shot translation capabilities, accurately translating between language pairs it had never seen before, such as Slovenian to Swahili.
Zero-shot learning allows AI systems to perform tasks or recognize objects without prior specific training. Traditional machine learning models require extensive data sets for each new task, but zero-shot learning algorithms apply existing knowledge to new situations and mimic human inferences.
The roots of zero-shot learning date back to a 2009 paper by researchers at Carnegie Mellon University titled “Zero-shot Learning with Semantic Output Codes.” Since then, the field has developed rapidly, with breakthroughs in the past five years.
From laboratory to real applications
The impact of zero-shot learning extends beyond language processing. Researchers published a study in Nature Biomedical Engineering demonstrating CheXzero, an AI model that detects various diseases from chest X-rays using zero-shot learning. The model successfully identified conditions it had never been trained on, such as pneumothorax (collapsed lung), by using its knowledge of related medical concepts and image features.
Google DeepMind unveiled Gato, a generalist AI agent that demonstrates zero-shot learning across domains. In one demonstration, Gato played a new Atari game he had never seen before, applying strategies he learned from other games to achieve a high score within minutes of exposure to the new game.
Zero-shot learning is also making waves in drug discovery. Researchers at MIT used a zero-shot learning model to predict the antimicrobial properties of molecules it had never encountered before. The model identified a new antibiotic compound that is effective against resistant bacteria, despite never having been trained in that specific class of antibiotics.
In the field of computer vision, Facebook AI (now part of Meta) developed Contrastive Language-Image Pre-training (CLIP), a zero-shot learning system. CLIP can classify images it has never seen before based solely on textual descriptions. For example, CLIP accurately categorized them using only their textual descriptions when presented with images of rare animals, and not in the training data.
Navigating challenges
Microsoft Research’s recent work on zero-shot learning in computer vision demonstrates models that identify objects in images without specific training. The system, VL-GPT, can generate detailed captions for images of complex scenarios it has never encountered before, such as describing the actions in a new sport or the components of an unknown technological device.
Robotics benefits from zero-shot learning. MIT’s Computer Science and Artificial Intelligence Laboratory showed robots manipulating previously invisible objects. In one experiment, a robot successfully grabbed and used twenty different tools that it had never encountered before, extrapolating its understanding of tool use to determine how to manipulate each new item.
Zero-shot learning represents a significant shift in AI development. The ability to generalize knowledge to new tasks could lead to more adaptable and efficient AI systems. Researchers at DeepMind predict that zero-shot learning could be critical to the development of artificial general intelligence (AGI), machines with human-like cognitive skills for a wide range of tasks.
As zero-shot learning advances, its impact on various industries is becoming increasingly apparent. From healthcare to finance, from robotics to language processing, this technology is poised to reshape the way we approach complex problems and interact with AI systems. In the coming years we will likely see an increase in the number of zero-shot learning applications, accompanied by ongoing debates about their implications for society, privacy and the future of work.
