Artificial Intelligence could provide a significant boost in the quest for a consolidated understanding of human cognition within the field of psychology.
In the ongoing quest to understand the complexities of human cognition, researchers are exploring the use of generative AI and large language models (LLMs) as potential tools to devise a unified theory.
These models, such as GPT-4, demonstrate high fluency and creativity in producing large volumes of ideas, making them valuable for exploring diverse cognitive mechanisms and generating novel theoretical insights or hypotheses rapidly. They can also serve as computational cognitive models, revealing how different cognitive capacities might emerge or interplay, offering a practical platform to test and refine unified cognitive theories.
However, these models are not without their limitations. They show a fixation bias, generating ideas within conventional categories and struggling to critically evaluate the originality or validity of ideas. This limitation means that human cognition is essential to assess, filter, and refine AI-generated hypotheses to avoid reinforcing existing cognitive biases or errors.
Moreover, the sociotechnical nature of AI-human interaction means that understanding cognition via AI requires focusing on AI as part of a cognitive relationship or system, not just an isolated model. This complicates the theoretical framing, as it necessitates considering the dynamic interplay between AI and human cognition.
Architecturally, existing transformer-based models are hitting theoretical limits in mimicking human cognition fully. This challenge may require new AI paradigms to better approximate unified cognitive functions.
Despite these challenges, the study introduced a foundation model of human cognition named Centaur, derived by fine-tuning a state-of-the-art language model on a large-scale dataset called Psych-101. The data augmentation included various types of psychology experiments, such as memory recall, supervised learning, decision-making, and others. The Centaur LLM was tested on hold-out data and showed excellent performance.
The researchers also tested Centaur LLM on out-of-distribution (OOD) settings, including economic games, deep sequential decision tasks, and reward learning, and it performed well in making predictions associated with those previously unseen experimental transcripts.
The mainstay of LLMs makes use of an artificial neural network (ANN), which is a series of mathematical functions that are computationally rendered in a computer system. The researchers made the dataset (Psych-101) and the augmented Meta Llama model (Centaur) available for other researchers to use and build upon.
The researchers chose to use Meta Llama as their base LLM and augmented it using the technique known as QLoRA (quantized low-rank adaptation). Some believe that the prevailing LLMs and AI architecture might not lead to artificial general intelligence (AGI) or artificial superintelligence (ASI).
The American Psychological Association (APA) defines cognition as "all forms of knowing and awareness, such as perceiving, conceiving, remembering, reasoning, judging, imagining, and problem solving." The enigma of human cognition remains a mystery, despite numerous attempts by researchers. The ongoing pursuit is to devise a unified theory of cognition.
In conclusion, while LLMs and generative AI introduce powerful new tools and perspectives that can accelerate and enrich research toward a unified theory of cognition, they are constrained by creativity biases, limited critical evaluation capacity, and the need for close human-AI collaboration. Overcoming these challenges involves addressing model limitations, developing hybrid human-AI frameworks, and rethinking cognitive theory to incorporate the dynamic sociotechnical AI-human interplay.
Artificial intelligence (AI), such as GPT-4, is being utilized in psychology and cognitive science to devise a unified theory of cognition, leveraging its high fluency and creativity to generate novel insights. Yet, the limitations of AI in critically evaluating ideas and avoiding cognitive biases necessitate human intervention in refining these hypotheses. (Science, technology, artificial-intelligence)
The development of a foundation model of human cognition, Centaur LLM, has showcased promising results in various cognitive tasks, although the ongoing quest is to make AI paradigms more sophisticated to better mimic unified cognitive functions, thereby approaching artificial general intelligence or artificial superintelligence. (Science, technology, artificial-intelligence)
