Agentic data science with the Data Agent Kit
cloud.google.com·6h ago
AI Research
arXiv CS.AI·Aug 7·5 min read
Small Foundation Models of Human Cognition and Behaviour
TL;DR
Fourteen small foundation models, ranging from 135M to 14B parameters, were trained on a dataset of human choices to evaluate their cognitive capabilities. Results indicate that smaller models can match larger ones in specific tasks, but larger models excel in generalizing to new task structures.
✦ Why It Matters
Consider using smaller models for specific cognitive tasks to reduce computational costs while maintaining performance.
Key Takeaways
- Explore using models with 0.6B to 1B parameters for cost-effective cognitive tasks.
- Evaluate model performance on new task structures to assess generalization capabilities.
- Utilize diagnostics to understand model reliance on specific input channels.
- Consider smaller models as noise ceiling estimators in psychological experiments.
How It Works
The study involved training models on a large dataset of human choices, allowing them to learn from various task structures. By progressively removing input channels, researchers could isolate the impact of different types of information on model performance, revealing the critical role of stimuli and feedback.