Centaur

1.

On DictofAI podcast, listen here!

https://spotifyanchor-web.app.link/e/AlWcgQqufOb

By erdal on Nov. 4, 2024, 11:32 a.m.

2.

a computational model trained to predict and simulate human behavior in a wide range of psychological experiments. Unlike previous models that specialized in specific cognitive domains, Centaur aims to capture the generality of human cognition, fulfilling a long-standing goal in cognitive science.
Here's a breakdown of Centaur's development and capabilities:

Centaur was built by fine-tuning a pre-existing, state-of-the-art language model called Llama 3.1 70B on a new dataset called Psych-101. The dataset is open on Github (see bellow)

The dataset contains trial-by-trial data from 160 psychological experiments across various domains, including decision-making, memory, learning, and reasoning. Notably, the data covers responses from over 60,000 participants performing over 10 million choices, making it an unprecedented resource in cognitive science.

A key feature of Psych-101 is the use of natural language to transcribe the experimental instructions and participant responses. This approach provides a consistent format for representing diverse experimental paradigms, facilitating the training of a domain-general model.

Fine-tuning: Researchers employed quantized low-rank adaptation (QLoRA), a parameter-efficient technique to fine-tune Llama on Psych-101. This method adds minimal trainable parameters to the base Llama model, preserving its knowledge while adapting it to human behavior.

Predictive Power: Centaur demonstrates superior performance in predicting human behavior compared to domain-specific cognitive models.

Generalization: Centaur exhibits remarkable generalization abilities, successfully predicting human behavior in:

Experiments with unseen participants,

Experiments with modified cover stories,

Experiments with modified problem structures,

Entirely novel domains that were not included in the training data

Despite being trained solely on behavioral data, Centaur's internal representations exhibit increased alignment with human neural activity, as evidenced by their ability to predict fMRI data.

The paper argues that Centaur is the first model to satisfy most of Newell's criteria for a unified theory of cognition, which include flexibility to respond to the environment, real-time operation, and use of vast knowledge about the world.

The researchers believe that Centaur represents a significant step towards a unified theory of human cognition, paving the way for future research exploring the inner workings of the mind.

Centaur's ability to predict and simulate human behavior has implications for the development of more human-like machines. As AI systems become increasingly integrated into our lives, it's crucial that they can interact with humans in a natural and intuitive way. By understanding human behavior at a deeper level, AI developers can create machines that are better equipped to collaborate, communicate, and adapt to human needs and preferences.

Paper (preprint): https://osf.io/preprints/psyarxiv/d6jeb

Github: https://huggingface.co/datasets/marcelbinz/Psych-101

By erdal on Nov. 4, 2024, 11:31 a.m.