Less Likely Brainstorming: Using Language Models to Generate Alternative Hypotheses | JoVE Visualize

Area of Science:

Artificial Intelligence
Cognitive Science
Medical Informatics

Background:

Human decision-making is prone to cognitive biases, limiting the utility of AI assistants focused solely on probable outcomes.
In fields like radiology, AI systems that only predict highly likely interpretations may not offer novel insights.
Addressing these limitations requires AI that can explore a broader range of possibilities, including less common ones.

Purpose of the Study:

To introduce and evaluate a novel AI task called "less likely brainstorming" designed to generate relevant but less probable outputs.
To develop and test a controlled text generation method that mitigates AI bias by considering a wider differential diagnosis.
To improve AI's ability to assist human decision-makers by generating diverse and less obvious interpretations.

Main Methods:

Exploration of the "less likely brainstorming" task in brain MRI interpretation and commonsense reasoning scenarios.
Development of a controlled text generation approach utilizing a novel contrastive learning strategy.
Comparison of the proposed method against baseline approaches and state-of-the-art controlled text generation models using automatic and human evaluations.

Main Results:

Standard Maximum Likelihood Estimation (MLE) training proved ineffective, with baseline models generating likely or irrelevant outputs frequently.
The proposed contrastive learning method demonstrated improved capability in generating less likely, yet relevant, outputs.
Human evaluations confirmed the enhanced performance of the new method in producing desired outputs.

Conclusions:

AI assistants can be enhanced to correct for human decision-making biases by generating less likely, relevant hypotheses.
The "less likely brainstorming" task and contrastive learning strategy represent a significant advancement in controlled text generation for AI.
This approach has the potential to improve AI's utility in complex decision-making domains, such as medical diagnosis and reasoning.