Confidence Minimization

Confidence minimization in machine learning aims to improve model reliability by reducing prediction confidence on uncertain or out-of-distribution inputs, thereby promoting safer and more robust AI systems. Current research focuses on techniques like test-time adaptation of large language and vision-language models, often employing low-rank updates to attention weights or Bayesian methods with calibration regularization. These approaches enhance both in-distribution accuracy and out-of-distribution detection, addressing critical needs in safety-critical applications and improving the trustworthiness of AI predictions. The ultimate goal is to develop more reliable AI systems that can accurately assess their own uncertainty and avoid making erroneous predictions in situations where they lack sufficient knowledge.

Papers

July 22, 2024

Test-Time Low Rank Adaptation via Confidence Maximization for Zero-Shot Generalization of Vision-Language Models
Raza Imam, Hanan Gani, Muhammad Huzaifa, Karthik Nandakumar
Zero Shot Vision Language Model Test Time Adaptation Pre Trained Vision Language Model Low Rank Adaptation Test Time Prompt Tuning Confidence Minimization

April 17, 2024

Calibrating Bayesian Learning via Regularization, Confidence Minimization, and Selective Inference
Jiayi Huang, Sangwoo Park, Osvaldo Simeone
Bayesian Learning Selective Inference Calibrated Learning Calibrated Out of Distribution Confidence Minimization

June 8, 2023

Conservative Prediction via Data-Driven Confidence Minimization
Caroline Choi, Fahim Tajwar, Yoonho Lee, Huaxiu Yao, Ananya Kumar, Chelsea Finn
Auxiliary Outlier Conservative Prediction Confidence Minimization

Confidence Minimization

Papers

Test-Time Low Rank Adaptation via Confidence Maximization for Zero-Shot Generalization of Vision-Language Models

Calibrating Bayesian Learning via Regularization, Confidence Minimization, and Selective Inference

Conservative Prediction via Data-Driven Confidence Minimization