Unifying Vision and Language Models with Mohit Bansal - 636

Bias Evaluation

Evaluating bias in generative AI models is a complex challenge that Mohit Bansal addresses through innovative metrics. He highlights the skewed distribution of gender and skin tone in models like DALL-Eval, emphasizing the need for metrics that correlate with human judgments to improve fairness 1. Sam Charrington adds that reinforcement learning with human feedback must consider diverse perspectives to avoid biases 2.

You always do want some human evaluation as a good complementary safeguard, but you also want to keep improving the metrics so that they correlate.

--- Mohit Bansal

This ongoing effort aims to ensure models can handle unseen scenarios without memorizing training data, moving towards more accurate and unbiased AI systems.

   

Explainability

Programmatic explainability in AI offers a promising approach to understanding model decisions. Mohit Bansal explains how generating programs that outline decision paths can enhance clarity and transparency in AI systems 3. These programs can integrate various APIs to perform diverse tasks, showcasing the potential for more comprehensive AI applications 4.

It gives you the whole program on how it came to the answer.

--- Mohit Bansal

This method not only provides answers but also illustrates the reasoning process, paving the way for more interpretable AI models.

   

Factuality

Ensuring factuality and faithfulness in multimodal models is crucial for maintaining accuracy. Mohit Bansal discusses the challenges of preventing hallucinations in AI outputs, emphasizing the importance of evaluating reasoning paths alongside final answers 5. He notes that multimodal models can leverage multiple data sources to improve faithfulness, though they also introduce complexity 6.

The model just getting the right answer is not enough.

--- Mohit Bansal

By developing metrics to assess both text and image alignment, researchers aim to enhance the reliability of AI-generated content.