Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Learning To Guide Human Decision Makers With Vision-Language Models

Debodeep Banerjee, Stefano Teso, Burcu Sayin, Andrea Passerini

TL;DR

This paper tackles the limitations of traditional hybrid decision making that relies on deferring uncertain cases to humans, which can lead to automation and anchoring biases. It introduces learning to guide (LTG), a framework where a machine generates interpretable textual guidance to assist humans, who retain full decision authority, and the slog method to convert vision-language models into task-specific guidance using limited human feedback. By training a surrogate model to estimate downstream decision quality and iteratively fine-tuning a VLM with an augmented loss that rewards guidance quality, slog demonstrates improved guidance informativeness and downstream decision accuracy on a real-world medical-imaging task. The findings suggest LTG enables safer, more reliable human-in-the-loop AI in high-stakes domains and point to future extensions via active learning and explainable-AI connections.

Abstract

There is increasing interest in developing AIs for assisting human decision-making in high-stakes tasks, such as medical diagnosis, for the purpose of improving decision quality and reducing cognitive strain. Mainstream approaches team up an expert with a machine learning model to which safer decisions are offloaded, thus letting the former focus on cases that demand their attention. his separation of responsibilities setup, however, is inadequate for high-stakes scenarios. On the one hand, the expert may end up over-relying on the machine's decisions due to anchoring bias, thus losing the human oversight that is increasingly being required by regulatory agencies to ensure trustworthy AI. On the other hand, the expert is left entirely unassisted on the (typically hardest) decisions on which the model abstained. As a remedy, we introduce learning to guide (LTG), an alternative framework in which - rather than taking control from the human expert - the machine provides guidance useful for decision making, and the human is entirely responsible for coming up with a decision. In order to ensure guidance is interpretable} and task-specific, we develop SLOG, an approach for turning any vision-language model into a capable generator of textual guidance by leveraging a modicum of human feedback. Our empirical evaluation highlights the promise of \method on a challenging, real-world medical diagnosis task.

Learning To Guide Human Decision Makers With Vision-Language Models

TL;DR

This paper tackles the limitations of traditional hybrid decision making that relies on deferring uncertain cases to humans, which can lead to automation and anchoring biases. It introduces learning to guide (LTG), a framework where a machine generates interpretable textual guidance to assist humans, who retain full decision authority, and the slog method to convert vision-language models into task-specific guidance using limited human feedback. By training a surrogate model to estimate downstream decision quality and iteratively fine-tuning a VLM with an augmented loss that rewards guidance quality, slog demonstrates improved guidance informativeness and downstream decision accuracy on a real-world medical-imaging task. The findings suggest LTG enables safer, more reliable human-in-the-loop AI in high-stakes domains and point to future extensions via active learning and explainable-AI connections.

Abstract

There is increasing interest in developing AIs for assisting human decision-making in high-stakes tasks, such as medical diagnosis, for the purpose of improving decision quality and reducing cognitive strain. Mainstream approaches team up an expert with a machine learning model to which safer decisions are offloaded, thus letting the former focus on cases that demand their attention. his separation of responsibilities setup, however, is inadequate for high-stakes scenarios. On the one hand, the expert may end up over-relying on the machine's decisions due to anchoring bias, thus losing the human oversight that is increasingly being required by regulatory agencies to ensure trustworthy AI. On the other hand, the expert is left entirely unassisted on the (typically hardest) decisions on which the model abstained. As a remedy, we introduce learning to guide (LTG), an alternative framework in which - rather than taking control from the human expert - the machine provides guidance useful for decision making, and the human is entirely responsible for coming up with a decision. In order to ensure guidance is interpretable} and task-specific, we develop SLOG, an approach for turning any vision-language model into a capable generator of textual guidance by leveraging a modicum of human feedback. Our empirical evaluation highlights the promise of \method on a challenging, real-world medical diagnosis task.
Paper Structure (14 sections, 3 equations, 4 figures, 4 tables)

This paper contains 14 sections, 3 equations, 4 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Hybrid Decision Making
  3. Beyond Separation of Responsibilities with Learning to Guide
  4. Estimating Downstream Decision Quality
  5. The slog Loop
  6. Benefits and Limitations
  7. Empirical Analysis
  8. Training the VLM
  9. Surrogate training
  10. Finetuning
  11. Results
  12. Related Work
  13. Conclusion
  14. Acknowledgments

Figures (4)

  • Figure 1: Left: Existing HDM algorithms employ a deferral function $d(\bm{\mathrm{x}})$ to partition the input space $\mathcal{X}$ into $\mathcal{H}$ and $\mathcal{M}$. Middle: A predictor $f(\bm{\mathrm{x}})$ handles the inputs falling in $\mathcal{M}$ (in blue). Because of anchoring bias, the human expert may end up blindly trusting its (possibly poor) decisions $y_m$. Right: The human, on the other hand, is left completely unassisted for those (possibly hard) decisions falling in $\mathcal{H}$, increasing the chance of mistakes in the human's decisions $y_h$ (in green).
  • Figure 2: The slog approach to learning to guide. Top: Given an input $\bm{\mathrm{x}}$, slog uses a VLM $\gamma\xspace$ to output textual guidance $g\xspace$ in support of human decision making. Here, $q$ indicates the quality of the human's downstream decision. Middle: The surrogate $\sigma_\mathrm{quality}$ estimates the quality of the downstream decisions and it is trained using a modicum on annotated guidance-quality pairs. Bottom: Given a trained surrogate $\sigma_\mathrm{quality}\xspace$, slog fine-tunes the VLM to output guidance $g$ achieving high (estimated) decision quality.
  • Figure 3: An example of a radiography along with its corresponding medical report consisting of 'findings' and 'impression'.
  • Figure 4: A qualitative example of the improvement of the guidance from slog with respect to the competitors. Green text indicates sentences that are (approximately) shared between the ground truth, slog and at least one of the competitors. Blue text indicates sentences shared between the ground truth findings (resp. impression) and slog, but missed by the competitors. No ground-truth sentences are shared between ground-truth text and competitors but missed by slog in this example.