QUTCC: Quantile Uncertainty Training and Conformal Calibration for Imaging Inverse Problems
summary
The gist
While deep learning offers tremendous promise for scientific and medical imaging, any failures and hallucinations (predictions that do not coincide with reality) are hard to pinpoint and can have…
In short
QUTCC uses a U-Net with quantile embeddings to learn a full conditional quantile distribution during training for imaging problems. It then uses this learned function for spatially-adaptive conformal calibration, producing tighter uncertainty intervals than previous methods. The method also predicts pixel-wise conditional probability density functions and can identify potential model hallucinations.
Key concepts
- Simultaneous Quantile Regression
- This technique trains a single neural network to predict any conditional quantile of the joint distribution by conditioning it on a learned 'quantile embedding.' This allows the model to learn the entire distribution shape rather than just fixed points, enabling non-linear adaptation during calibration.
- Spatially-Adaptive Conformal Calibration
- Instead of calibrating a single scaling parameter, QUTCC finds network parameters that guarantee coverage on a pixel-by-pixel basis. This spatial adaptation leverages the learned quantile function to produce uncertainty intervals that are tighter and more accurate across different parts of an image compared to traditional methods.
- Pixel-wise Conditional PDF Estimation
- The method estimates the conditional probability density function, p(x|y), by querying the quantile-conditioned network multiple times across a range of quantile levels. Numerical differentiation of this resulting quantile function provides an estimate of the underlying distribution without needing to assume any specific statistical model.
- Quantile Embedding
- A continuous scalar embedding is used to condition the U-Net's self-attention layers. This embedding represents a target quantile level, allowing the network to adapt its predictions based on where it needs more or less uncertainty information within the distribution during training.
Terminology used across episodes
This episode discusses
- QUTCC: Quantile Uncertainty Training and Conformal Calibration for Imaging Inverse Problems · Paper Radio
- fastMRI: An Open Dataset and Benchmarks for Accelerated MRI
- A Gentle Introduction to Conformal Prediction and Distribution-Free Uncertainty Quantification
- An Information Theoretic Perspective on Conformal Prediction
- Uncertainty Sets for Image Classifiers using Conformal Prediction
- Conformal Language Modeling
- Attention U-Net: Learning Where to Look for the Pancreas
- Adam: A Method for Stochastic Optimization
- The Berkeley Single Cell Computational Microscopy (BSCCM) Dataset
- On the Posterior Distribution in Denoising: Application to Uncertainty Quantification
- Bayesian Convolutional Neural Networks with Bernoulli Approximate Variational Inference
The paper
QUTCC: Quantile Uncertainty Training and Conformal Calibration for Imaging Inverse Problems · Read on arXiv
While deep learning offers tremendous promise for scientific and medical imaging, any failures and hallucinations (predictions that do not coincide with reality) are hard to pinpoint and can have serious downstream consequences. Uncertainty estimation techniques, such as conformal prediction, can help by predicting statistically valid error bars for a model's prediction. However, popular conformal prediction methods were not designed for high-dimensional image-valued problems and do not take into account spatial correlations within an image during conformal calibration, resulting in larger-than-necessary uncertainty intervals. We propose a practical simultaneous quantile regression method that enables non-linear, spatially-adaptive scaling during conformal calibration. Our method, QUTCC uses a U-Net architecture with a quantile embedding to learn a full conditional quantile distribution during training, and then leverages this non-linear, learned function for spatially-adaptive conformal calibration. At test time, our method can efficiently estimate uncertainty intervals with pixel-marginal coverage guarantees. In addition, QUTCC can also predict pixel-wise conditional probability density estimates without any built-in distributional assumptions. We evaluate our method on several denoising problems, accelerated magnetic resonance imaging, and quantitative phase microscopy. Our method consistently produces tighter uncertainty intervals than prior conformal methods at the same coverage level, can predict plausible conditional distributions for different tasks, and in some cases, high-uncertainty regions can help us locate hallucinations in a model's prediction.
Transcript
Introduction to the show: ident: Electrical Engineering. The day in electrical engineering: signals, systems and the hardware that carries them.
Rosa: We're looking at the paper: "QUTCC: Quantile Uncertainty Training and Conformal Calibration for Imaging Inverse Problems".
Title and authors: Dev: Now, let's look at the title of "QUTCC: Quantile Uncertainty Training and Conformal Calibration for Imaging Inverse Problems." It tells us immediately that it’s focused on two main things: quantile uncertainty training and conformal calibration for imaging problems.
Rosa: That's right. The authors are Cassandra Tong Ye, Shamus Li, Tyler King, and Kristina Monakhova. They tackle the issue of how to get uncertainty in image reconstruction when you don't have perfect ground truth data readily available for every pixel.
Taro: What does that mean practically? It means instead of getting a single error bar for the whole image, they can potentially tell you which specific parts of the image are most uncertain.
Dev: That’s what it suggests. It moves away from just giving a general confidence score to providing a pixel-wise uncertainty map based on how well the model is conditioned on different quantile levels.
Rosa: They propose using a U-Net architecture with this quantile embedding during training, and then they leverage that learned function for spatially-adaptive conformal calibration, which results in tighter uncertainty intervals than prior conformal methods at the same coverage level.
Taro: Tighter intervals mean less false confidence when you're making a decision based on that image.
Dev: And they also introduce a way to estimate and calibrate a pixel-wise conditional probability density function by querying that same quantile-conditioned network over several quantile levels and using numerical differentiation to reconstruct the distribution.
Rosa: So, beyond just error bars, they’re predicting the actual shape of the probability distribution for what's happening at each pixel.
Taro: That’s useful because it gives you a richer picture of what might be wrong with your reconstruction than just a single upper and lower bound.
The paper's summary: Rosa: So, to sum up the core of this paper, QUTCC builds on conformal prediction but adds a practical simultaneous quantile regression method that lets the network learn the full conditional quantile distribution during training using an asymmetric pinball loss.
Dev: The key training signal they use is not just predicting a single image, but predicting any conditional quantile q of the joint distribution by conditioning the neural network on a continuous scalar called a quantile embedding.
Taro: That sounds complex for real-time systems. How does that learned function actually help in practice when we're running things live?
Rosa: The architecture is a U-Net backbone augmented with self-attention mechanisms where those quantile embeddings are propagated through the attention layers to guide the network’s predictions of the quantile.
Dev: So, during calibration, instead of just tuning some scaling parameter directly, they find the neural network conditioning parameters called qlo and qhi that achieve your target coverage level alpha, enforcing marginal pixel-wise coverage guarantees using two conditions based on the ground truth.
Taro: That’s where I see a potential failure mode. If those bounds are set too high or too low, the calibration might not be robust enough for the real world data we throw at it.
Rosa: They address that by proposing this spatial adaptation, which is what enables non-linear scaling during conformal calibration instead of relying on fixed procedures as in older methods like Im2Im or K-RCPS.
Dev: And they then use this trained network to predict the mean image, a pixel-wise uncertainty map, or even a conditional quantile distribution when you query it at test time.
Taro: So, if I'm looking at an output and the uncertainty map shows high uncertainty right where my object is supposed to be, that tells me something specific about the model's failure mode.
The paper's improvements: Rosa: The main improvement they highlight is that this quantile regression with quantile embedding allows a single neural network to learn a full conditional quantile distribution during training, which enables non-linear, spatially-adaptive scaling during conformal calibration.
Dev: That’s the big win. It means the method can produce tighter uncertainty intervals than prior conformal methods when you're aiming for the same coverage level alpha.
Taro: Tighter uncertainty intervals are better for risk assessment in any application, whether it's autonomous driving or medical diagnostics.
Rosa: They also introduce a method to estimate and calibrate a pixel-wise conditional probability density function p(xy) by querying the quantile-conditioned network over a range of quantile levels and using numerical differentiation to estimate the derivative.
Dev: That’s important because it lets you reconstruct that underlying distribution without any built-in distributional assumptions, which is something many other methods rely on.
Taro: If we can get that distribution estimate without those assumptions, it makes the method more general across different types of data we might encounter in the field.
Rosa: The result is they can predict plausible conditional distributions for different tasks, and they show this works by demonstrating that high uncertainty regions are co-located with a high reconstruction error, which helps locate model hallucinations.
Conclusion: Dev: So to wrap up, the QUTCC paper shows how quantile regression with quantile embeddings allows a single neural network to learn the full conditional quantile distribution during training, enabling non-linear scaling during conformal calibration.
Rosa: And by using this learned function for spatially-adaptive conformal calibration, they achieve tighter uncertainty intervals and provide a way to predict pixel-wise conditional probability density functions without distributional assumptions.
Taro: It seems like the ability to pinpoint high uncertainty regions where failures or hallucinations are likely is a really practical tool for testing and deploying these models in the real world.
Dev: That's the implication: when you're using deep learning for imaging inverse problems, you can move beyond just getting an error bar and start understanding the shape of what could be wrong at each location.
Rosa: It’s a lot to unpack, but QUTCC gives us a way to build more trustworthy uncertainty estimation for these complex tasks.
Taro: I'm still curious how robust that spatial adaptation is when the world misbehaves and the data is noisy in unexpected ways outside of their training set.
Dev: That’s exactly what we need to keep asking about as we move toward real-time loops, because having a safety case for that non-linear scaling is something engineers have to worry about.
Rosa: Right, so QUTCC gives us a framework to build more trustworthy uncertainty estimation for these complex tasks.
More episodes
- Daily Summary for 2026-11-19
- 2111.07910-Mask-guided Spectral-wise Transformer for Efficient Hyperspectral Image Reconstruction
- 2203.02149-HDNet: High-resolution Dual-domain Learning for Spectral Compressive Imaging
- 2205.10102-Degradation-Aware Unfolding Half-Shuffle Transformer for Spectral Compressive Imaging
- 2311.10959-Structure-Aware Sparse-View X-ray 3D Reconstruction
- 2403.04116-Radiative Gaussian Splatting for Efficient X-ray Novel View Synthesis
- 2410.02572-Combining Pre- and Post-Demosaicking Noise Removal for RAW Video
- 2604.17453-Learned Nonlocal Feature Matching and Filtering for RAW Image Denoising
- 2603.20387-End-to-End Multi-Task Learning for Adjustable Joint Noise Reduction and Hearing Loss Compensation
- 2507.20825-Chirp-Permuted AFDM: A Versatile and Multifunctional Waveform for 6G