What’s zero shot learning performance in AI?
Zero-shot learning performance in artificial intelligence (AI) and cognitive automation is getting more scrutiny. So, I thought I’d take time to look at what zero-shot learning is, as well as review research on it. To understand zero-shot learning, however, we have to understand machine learning in AI.
Machine learning in AI
Machine learning (ML) works by following instructions from algorithms, programmed to detect patterns among data. Those patterns can be used to form generalizations about the information. And those generalizations can mimic how humans learn. For example, the data might be various photos of Poodles. ML would find patterns among these photos. Then, it would use those patterns to form generalizations about what Poodles typically look like.
Of course, these generalizations (about what Poodles look like) are based on specific examples (of Poodles in those photos). And those examples come from a limited set of data. For this reason, the generalizations may also be limited, so they aren’t always perfect, particularly when applied beyond that data set—an idea that we’ll circle back to.
Still, once ML identifies patterns in enough Poodle pictures, it may be able to analyze other similar photos and correctly classify which ones are of Poodles. In this way, the information patterns—and the generalizations that may come out of them—can imitate human learning.
Hence, the term “machine learning”: algorithms appear to learn by finding patterns in data and forming generalizations about that information.
Zero-shot learning performance in ML
Zero-shot learning is a capability being pursued in ML. The goal is for AI to perform a new task without having learned from any prior examples of that task. For example, the AI may try to identify photos of Poodles, without having classified any Poodle pictures beforehand.
How’s that possible? In general, zero-shot learning performance may involve different methods that help AI make informed guesses about tasks, based on information that it already learned from similar tasks. For instance, the AI may just have information that Poodles are a kind of dog with thick, curly fur. By classifying which dogs have thick, curly fur, it can identify which ones are Poodles.
.jpg)
Is zero-shot learning performance in AI a long shot?
Remember: ML works from a limited sample of examples. This limited data set means that the generalizations coming out of the information patterns may have their limits too. Given such limitations, a perfect zero-shot learning performance in AI may remain a long shot. Indeed, research shows that, for ML to continually improve zero-shot performance, it may require exponentially more data, especially as its scale increases.
So, it seems zero-shot learning can run into an age-old philosophical dilemma: the problem of induction. Generalizing beyond a limited set of data may lead to errors, such as false or misleading conclusions. It’s yet another reason why hallucinations and biases in AI are unlikely to go away anytime soon.
Related posts
AI chatbots: artificial general intelligence or cognitive automation?
Technological innovation and the future of work: synopsis of ‘The Second Machine Age’
Jailbreaking: a feature, not a bug, of general-purpose chatbots
Note: For the record, I’m a UX and technical communication professional who has worked in the conversational AI space. But I’m not a programmer or computer scientist. So, if you think I missed something about the realizability of zero-shot learning, feel free to share those thoughts in a comment below.