Neural Network Achieves Human-Like Language Generalization, Surpassing ChatGPT

Scientists have created a neural network that can generalize language in the same way that humans can, which is a significant step forward in the field of artificial intelligence. This AI can quickly learn and incorporate new words into its lexicon for several contexts. Systematic generalization is an essential component of human intellect. 

The prominent journal Nature recently recognized the work’s success in this sector. Surprisingly, it outperformed the AI model employed by the popular ChatGPT chatbot. Despite great praise for its human-like conversational abilities, ChatGPT could have done better when put through a thorough generalization test; however, the freshly constructed neural network performed on par with human volunteers. 

The human brain is designed to integrate new terms into various settings effortlessly. A person who understands the concept of “photobomb” may readily use it in a variety of contexts, such as “photobomb during a meeting” or “photobomb twice.” This potential for lexical modification and transfer to novel situations exemplifies the human brain’s plasticity and adaptability.

Contrary to common assumptions, neural networks do not have an intrinsic aptitude for humans’ cognitive flexibility. Effective use of a novel word may involve extensive training of artificial intelligence models to replicate human cognition. For years, artificial intelligence researchers have disputed this limitation. Given their apparent shortcomings, the debate centers on whether neural networks can serve as an accurate model for human cognition. 

The research team wanted to understand more about this debate and provide genuine facts, so they conducted a specific experiment with 25 participants. They taught the members of the organization a false language. This made-up language has two kinds of words. Words like ‘dax’ and ‘lug’ belong to the first category, known as ‘primitive’ words.

These words signified simple actions such as “jump” and “skip.” Terminologies like ‘blicket’ and ‘fep’ belong to the second group, known as ‘function’ terminology. These function words served as a criterion for combining the fundamental words, resulting in phrases such as “jump three times” or “skip backward.” Then, students had to match each simple word with the corresponding colored circle. ‘dax’ may be represented by a red circle, whereas ‘lug’ may be represented by a blue circle.

Participants were then shown combinations of primitive and function words supplied by the researchers. In addition to these combinations, they displayed the outcomes of applying the functions to the primitives as patterns of circles. The main challenge for participants was interpreting and applying general standards to odd combinations of these sentences. 

Humans did wonderfully in this experiment, as predicted. Their success rate was close to 80%, which is excellent. When people did make mistakes, they frequently followed predictable patterns. These patterns corresponded to well-known human cognitive biases, illuminating the mind’s intricacies. 

The scientists next went about training a neural network. They used a dynamic method that included real-time AI input to increase its performance. This method differed from the standard practice of employing static dataset training in AI research. The AI was educated to emulate the error patterns identified in human test results, which made its learning more human-centric.

Further experiments on this neural network produced findings quite close to those obtained with human volunteers. It even outperformed humans in numerous situations, illustrating how far it may go. GPT-4, an alternate AI model, had a considerably more difficult time. Error rates ranged from 42% to 86%, depending on how the job was presented. The significant disparity in results between the two AI models was informative regarding the nuances of AI training. 

Brenden Lake, one of the study’s authors, emphasized the need for experience while teaching AI. He likened AI learning to that of a baby learning a new language. He found that regular exposure to new compositional challenges refines and hones their skills, like a child who learns and adapts over time. 

Elia Bruni, a natural language processing expert, explored the prospects of adding structure to neural networks. Such advancements have the potential to significantly reduce the vast volumes of data now necessary for training complex AI systems like ChatGPT. The issue of ‘hallucination,’ in which AI identifies patterns that do not exist, may also be addressed. 

Finally, the results of this groundbreaking study lead to a better future for artificial intelligence. It emphasizes the potential benefits of including human-like systematic generalization in AI systems. By reducing the gap between human cognition and machine learning, scientists are paving the way for more human-like AI systems.  

Journal Reference  

Kozlov, M., & Biever, C. (2023). AI “breakthrough”: neural net has human-like ability to generalize language. Retrieved from https://www.nature.com/articles/d41586-023-03272-3