Centaur: az AI modell, amely úgy gondolkodik, mint egy ember?

A computational model called Centaur, developed by fine-tuning a language model on a huge dataset called Psych-101, can predict and simulate human nature in experiments expressible in natural language, even in previously unseen situations.

Helmholtz Munich, is a member of the Helmholtz Association of German Research Centres

The paper presents the Centaur model, a novel computational framework developed to serve as a foundation model of human cognition. The key objective of this model is to predict and simulate human behavior across various psychological experiments, through a data-driven methodology that finetunes a state-of-the-art language model on a domain-specific dataset known as Psych-101. This dataset comprises data from 160 psychological experiments, with over 60,000 participants providing more than 10 million choices, creating a robust structure for evaluating human cognition. **Core Methodology:** 1. **Data Collection**: - The dataset, Psych-101, encompasses multiple psychological domains such as multi-armed bandits, memory tasks, decision-making, and more. Each experiment is carefully transcribed into natural language, enabling uniformity across different experimental setups. - The inclusion criteria for experiments focused on availability, the capacity for transcription without significant information loss, and a diverse range of cognitive domains. 2. **Model Architecture**: - Centaur is built upon the Llama 3.1 70B language model, a powerful pre-trained model developed by Meta AI. The model leverages the knowledge embedded in Llama to predict human responses effectively. - The paper employs Quantized Low-Rank Adaptation (QLoRA), a parameter-efficient finetuning technique that attaches low-rank adapters to the model’s layers while keeping the base model’s parameters unchanged. The low-rank adapters contain only a fraction of the model’s parameters (0.15%), focusing on optimizing task-specific learning while maintaining the base model’s integrity. 3. **Finetuning Procedure**: - Centaur is finetuned on the Psych-101 dataset using a supervised learning approach characterized by cross-entropy loss computations. This step ensures the model prioritizes human behavioral data over other input noise. - To enhance efficiency and avoid overfitting, the model is trained for one epoch across the entire dataset, with gradients only backpropagated for tokens corresponding to human responses. 4. **Evaluation Metrics**: - The effectiveness of the Centaur model is measured via Pseudo-R² metrics, where predictions from Centaur can be benchmarked against human responses. This metric normalizes model log-likelihoods and provides a uniform scale across experimental categories. - Evaluations also include comparisons with domain-specific cognitive models, demonstrating Centaur’s superior performance across various tasks, often exceeding the predictive capabilities of other models. 5. **Generalization Capabilities**: - The model’s ability to generalize is assessed through a series of rigorous tests. These tests include variations in cover stories (e.g., changing the narrative context of a decision-making task) and task structures (e.g., modifying a multi-armed bandit task). - Centaur’s ability to predict responses in settings outside its original training data showcases distinct robustness and transfer learning capabilities. 6. **Behavioral Simulations**: - The model is subjected to open-loop simulations to examine its predictive ability concerning human-like behavior generated from its own outputs. Simulations in various experimental paradigms confirm that Centaur can replicate patterns similar to those of the human population. 7. **Neural Alignment**: - An important aspect investigated in the paper is whether the internal representations within Centaur become aligned with human neural responses. Studies using fMRI data confirm that Centaur’s representations not only improve in predictability following finetuning but also show significant correlation with neural activity patterns observed in human subjects. 8. **Future Directions**: - The authors intend to expand the dataset to include additional domains of cognition, reflecting a broader array of human behaviors. They plan to leverage tools for exploring knowledge representation and processing within cognitive frameworks, ultimately hoping to construct a unified theory of human cognition based on models like Centaur. In conclusion, Centaur is a pioneering approach towards a unified model of human cognition, integrating advanced machine learning techniques with expansive psychological data. It not only leverages vast datasets for enhancing predictive accuracy but also shows promise in understanding the underlying cognitive processes that characterize human thought and decision-making. The implications of this research could redefine methodologies within cognitive sciences and inform the development of future models.