# Predictive Inference with Feature Conformal Prediction

**Authors:** Jiaye Teng, Chuan Wen, Dinghuai Zhang, Yoshua Bengio, Yang Gao, Yang Yuan  
**arXiv:** [2210.00173](https://arxiv.org/abs/2210.00173)  
**Date:** 2022-10-01  
**Summary page:** https://www.summarizepaper.com/en/paper/2210.00173v4/

## Key points

- Introduces a technique called feature conformal prediction
- Expands the scope of conformal prediction to semantic feature spaces
- Conducts conformal prediction in semantic feature spaces using deep representation learning
- Outperforms regular conformal prediction from a theoretical perspective
- Can be combined with other adaptive conformal prediction methods
- Showcases state-of-the-art performance on large-scale tasks such as ImageNet classification and Cityscapes image segmentation
- Provides code for implementation at a GitHub repository for public use
- Authors: Jiaye Teng, Chuan Wen, Dinghuai Zhang, Yoshua Bengio, Yang Gao, and Yang Yuan
- Published as a conference paper at ICLR 2023 under categories cs.LG, cs.AI, stat.ME, and stat.ML

## Summary

The paper titled "Predictive Inference with Feature Conformal Prediction" introduces a novel technique called feature conformal prediction, which expands the scope of conformal prediction to semantic feature spaces. Conformal prediction is a distribution-free method used to establish valid prediction intervals. While traditional conformal prediction is conducted in the output space, this paper proposes conducting it in semantic feature spaces by leveraging the inductive bias of deep representation learning. The authors demonstrate that feature conformal prediction outperforms regular conformal prediction under mild assumptions from a theoretical perspective. This approach can be combined not only with vanilla conformal prediction but also with other adaptive conformal prediction methods. In addition to experiments on existing predictive inference benchmarks, the authors showcase the state-of-the-art performance of their proposed methods on large-scale tasks such as ImageNet classification and Cityscapes image segmentation. Furthermore, they provide code for implementing these methods at a GitHub repository for public use. The authors of this paper are Jiaye Teng, Chuan Wen, Dinghuai Zhang, Yoshua Bengio, Yang Gao and Yang Yuan and was published as a conference paper at ICLR 2023 falling under the categories of cs.LG, cs.AI, stat.ME and stat.ML.