Paper Image

Making machine learning models more human-compatible: Key insights from a study on image classification

Published on:

6 June 2023

Primary Category:

Machine Learning

Paper Authors:

Eleni Straitouri,

Manuel Gomez Rodriguez

Bullets

Key Details

Conducted large-scale study with 2,751 participants classifying images using machine learning model predictions

Forcing model to predict from small set of options increased human accuracy vs free choice

Assumption of user behavior let algorithm quickly hone in on optimal prediction set size

Strict system limiting user agency outperformed more lenient system

Findings suggest limiting model autonomy can improve human-AI collaboration

AI generated summary

Making machine learning models more human-compatible: Key insights from a study on image classification

This paper presents a human subject study evaluating different strategies for improving collaboration between humans and machine learning models on an image classification task. The study finds that limiting the model's predictions to a small 'prediction set' leads to higher accuracy compared to giving the model free reign. The paper also shows that modeling user behavior and adapting the prediction set accordingly further improves performance. Overall, the results suggest that constraining model autonomy based on user capabilities is a promising approach for human-AI collaboration.

Answers from this paper

Comments

No comments yet, be the first to start the conversation...

Sign up to comment on this paper

Sign Up