Robustness and Exploration of Variational and Machine Learning Approaches to Inverse Problems: An Overview

Alexander Auras, Kanchana Vaishnavi Gandikota, Hannah Droege, Michael Moeller

TL;DR

This review explores the critical tension between reconstruction quality and stability in solving imaging inverse problems. By comparing classical variational methods with modern deep learning, the authors highlight a fundamental trade-off: while neural networks produce visually superior results, they are often significantly more vulnerable to adversarial noise than their mathematically grounded predecessors.

What Problem Does This Paper Address?

Inverse problems involve recovering a hidden signal (like a clean image) from corrupted or indirect measurements (like a blurry photo or a raw MRI scan). These problems are often "ill-posed," meaning small amounts of noise in the data can lead to massive, unpredictable errors in the reconstruction.

While deep learning (DL) has set new records for reconstruction accuracy, these models are often "black boxes" lacking the stability guarantees of classical mathematics. This paper investigates the robustness of these methods—specifically how they handle adversarial attacks designed to fool them—and how we can "explore" the multiple valid solutions that exist when a problem is underdetermined.

Methods

The authors categorize current approaches into two main camps:

  1. Point Estimators: Methods that aim to provide a single "best guess" solution.
  2. Variational Methods: The classical approach, using a cost function that balances data fidelity with a "regularizer" (prior knowledge). Examples include Tikhonov regularization and Total Variation (TV).
  3. Learned Approaches: These range from fully learned "black box" networks (like U-Nets) to "unrolled" optimization schemes that turn iterative math algorithms into neural network layers.
  4. Neural Network Priors: Using trained networks as sophisticated regularizers (e.g., Plug-and-Play methods) or even using untrained networks as a structural prior (Deep Image Prior).
  5. Bayesian & Generative Approaches: Instead of one guess, these model the entire probability distribution of possible solutions. This includes GANs, VAEs, and more recently, Diffusion Models.

The paper also introduces Explorability, using guidance (like text prompts or classifiers) to navigate the space of possible solutions to find one that meets specific semantic criteria.

Datasets

To provide a fair comparison, the authors conducted numerical experiments on a synthetic 1D toy problem:

  • Signal: Piece-wise constant functions with a random number of jumps (simulating simplified image edges).
  • Task: Compressed sensing—recovering the signal from a significantly smaller set of random measurements.
  • Training: Learned models were trained on a dataset of 8192 samples with added white Gaussian noise.

Key Insights and Ideas

  • The Robustness-Fidelity Trade-off: There is a clear "no free lunch" theorem at play. Models that achieve the lowest error on standard noisy data are often the most fragile when faced with adversarial perturbations.
  • The Flaw of Convexity: Classical convex regularizers are mathematically convenient but physically limited. As shown in Figure 2, a convex regularizer considers a blurry average of two natural images to be "more natural" than the original images themselves. Convex regularizers rating convex combinations of images
  • Stability Metrics: While $L_2$ (mean squared error) stability is the standard in ML, variational methods offer stability in terms of the Bregman distance. This metric is more natural for inverse problems because it accounts for the specific regularizer being used.
  • Explorability via Guidance: In highly underdetermined problems (like 16x super-resolution), there are many "correct" solutions. The authors show how text-to-image diffusion models can be guided to produce different, equally valid reconstructions based on user intent (e.g., "a smiling woman" vs. "a man").
Convex regularizers rating convex combinations of images

Key Results

The authors compared several models on their 1D task, including Tikhonov, TV, a post-processing U-Net, and ItNet (an iterative learned network).

  • Top Performer: ItNet achieved the best reconstruction quality on standard noisy data ($L_2$ error of 0.34).
  • The Catch: Under a targeted adversarial attack, ItNet's error skyrocketed to 23.57. In contrast, a highly regularized Tikhonov model was much more stable, though its initial reconstruction was far blurrier.
  • Bound Verification: The authors successfully verified theoretical stability bounds for variational methods. As seen in Figure 4, while TV reconstructions (e) are visually closer to the ground truth than Tikhonov (a), they respond differently to adversarial noise (f vs b). 1D signal reconstruction results
1D signal reconstruction results
Method Noisy Input Error ($L_2$) Adversarial Error ($L_2$)
Tikhonov ($\alpha=10^{-7}$) 4.39 86.37
ItNet 0.34 23.57
Total Variation (TV) 0.66 11.54

Limitations and Future Work

  • Computational Complexity: Proving formal stability (Lipschitz constants) for deep networks remains NP-hard, making it difficult to provide the "guarantees" required for medical or safety-critical applications.
  • Model Adaptivity: End-to-end trained networks are often brittle; if the physics of the measurement (the forward operator $A$) changes slightly, the network usually needs to be entirely retrained.
  • Future Direction: The authors suggest that hybrid models—combining the interpretability and stability of variational methods with the expressive power of diffusion models—represent the most promising path forward for robust image reconstruction.

Summary updated

Abstract

This paper provides an overview of current approaches for solving inverse problems in imaging using variational methods and machine learning. A special focus lies on point estimators and their robustness against adversarial perturbations. In this context results of numerical experiments for a one-dimensional toy problem are provided, showing the robustness of different approaches and empirically verifying theoretical guarantees. Another focus of this review is the exploration of the subspace of data-consistent solutions through explicit guidance to satisfy specific semantic or textural properties.
BibTeX
@misc{auras2024robustness,
  author = {Alexander Auras and Kanchana Vaishnavi Gandikota and Hannah Droege and Michael Moeller},
  title = {{Robustness and Exploration of Variational and Machine Learning Approaches to Inverse Problems: An Overview}},
  year = {2024},
  eprint = {2402.12072},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2402.12072},
}
Figure