SPC-UQ: A Post-hoc, Efficient, and Unified Uncertainty Quantification Framework

This repository contains the official code for SPC-UQ (Split-Point Consistency for Uncertainty Quantification), a post-hoc framework that jointly quantifies aleatoric and epistemic uncertainty with a single forward pass.

It accompanies the paper:

"Post-Hoc Split-Point Self-Consistency Verification for Efficient, Unified Quantification of Aleatoric and Epistemic Uncertainty in Deep Learning."

Key Features

• Post-hoc – augments pre-trained network without architectural changes and retraining.
• Unified – supports both regression and classification tasks in deep learning.
• Efficient – produces aleatoric and epistemic uncertainty estimates in one forward pass.
• Calibration – provides mechanisms to calibrate aleatoric uncertainty and improve predictive reliability.

Repository Structure

Cubic_Regression/             # Toy cubic regression for fast demonstration.
MNIST_Classification/         # Digit classification for fast demonstration.
UCI_Benchmarks/               # Standard UCI regression datasets for scalar regression evaluation.
Monocular_Depth_Estimation/   # Monocular end-to-end image depth estimation for high-dimensional regression.
Image_Classification/         # CIFAR-10/100, ImageNet-1K for large-scale image classification.
Multimodal_Classification/    # LUMA multimodal benchmark (image/audio/text) for multimodal classification tasks.

Each directory provides scripts to reproduce the corresponding experiments.

Installation

We recommend using conda to manage dependencies.
All required packages and versions except Multimodal_Classification are specified in environment.yml.

Step 1: Clone the repository

huggingface-cli download zzz0527/SPC-UQ --local-dir SPC-UQ
cd SPC-UQ

Step 2: Create and activate the environment

conda env create -f environment.yml
conda activate spc_uq

Usage

Each subdirectory corresponds to a specific benchmark.
To run an experiment, navigate into the corresponding folder and follow the instructions provided in its README.md.

Quick Start

For a fast verification, we provide two lightweight benchmark tasks:

# Synthetic cubic regression
python Cubic_Regression/run_cubic_tests.py

# MNIST classification
python MNIST_Classification/run_cls_tests.py

See the documentation in each subdirectory for details on dataset preparation, configuration options, and advanced usage.

Citation

If you use SPC-UQ in your research, please cite our paper:

@article{zhao2025post,
  title={Post-Hoc Split-Point Self-Consistency Verification for Efficient, Unified Quantification of Aleatoric and Epistemic Uncertainty in Deep Learning},
  author={Zhao, Zhizhong and Chen, Ke},
  journal={arXiv preprint arXiv:2509.13262},
  year={2025}
}