Abstract
Methods inspired by Artificial Intelligence (AI) are starting to fundamentally change computational science and engineering through breakthrough performances on challenging problems. However, reliability and trustworthiness of such techniques is becoming a major concern. In inverse problems in imaging, the focus of this paper, there is increasing empirical evidence that methods may suffer from hallucinations, i.e., false, but realistic-looking artifacts; instability, i.e., sensitivity to perturbations in the data; and unpredictable generalization, i.e., excellent performance on some images, but significant deterioration on others. This paper presents a theoretical foundation for these phenomena. We give a mathematical framework describing how and when such effects arise in arbitrary reconstruction methods, not just AI-inspired techniques. Several of our results take the form of 'no free lunch' theorems. Specifically, we show that (i) methods that overperform on a single image can wrongly transfer details from one image to another, creating a hallucination, (ii) methods that overperform on two or more images can hallucinate or be unstable, (iii) optimizing the accuracy-stability trade-off is generally difficult, (iv) hallucinations and instabilities, if they occur, are not rare events, and may be encouraged by standard training, (v) it may be impossible to construct optimal reconstruction maps for certain problems, (vi) standard methods to improve reliability (e.g., regularization or adversarial training) may themselves lead to unstable problems. Our results trace these effects to the kernel of the forwards operator. They assert that such effects can be avoided only if information about the kernel is encoded into the reconstruction procedure. Based on this, this work aims to spur research into new ways to develop robust and reliable AI-inspired methods for inverse problems in imaging.
Introduction
In imaging science, we frequently encounter inverse problems: the task of reconstructing a high-quality image $x$ from imperfect, noisy, or undersampled measurements $y$. While Deep Learning (DL) has achieved record-breaking performance in these tasks, it suffers from three major "trustworthiness" issues: hallucinations (realistic-looking artifacts), instability (sensitivity to noise), and unpredictable generalization.
This paper provides a rigorous mathematical foundation for these failures, proving that they are not just training glitches but fundamental "no free lunch" trade-offs between accuracy and stability inherent in the physics of inverse problems.
The Physics of the Kernel
The core of the issue lies in the kernel (or null space) of the forward operator $A$—the mathematical description of the imaging hardware. The kernel represents the "blind spots" of the sensor: details that exist in the real world but produce zero signal in the measurements. If a neural network "recovers" a detail that lies in the kernel, it is not "sensing" it; it is "inventing" it based on its training data. This invention is the root cause of hallucinations.
The Accuracy-Stability Trade-off
The authors establish several "no free lunch" theorems for reconstruction maps. A central finding is a fundamental limit: as you push a model to be more accurate on a specific set of images (driving training error to zero), it must eventually become unstable. In ill-conditioned problems, perfect accuracy and perfect stability are mutually exclusive.
Furthermore, standard training practices that reward low pixel-wise error (like Mean Squared Error) actually encourage the model to hallucinate details from the training set to "fill in the blanks" left by the sensor's blind spots.
Hallucinations and Detail Transfer
Theorem 4.1 (Detail Transfer) proves that any map that successfully recovers a "hidden" detail in one image will necessarily hallucinate that same detail when reconstructing a "detail-free" image. The authors validated this using "adversarial" setups where recognizable shapes, such as a Mickey Mouse silhouette, were added to training data.
As shown in Figure 6, a model trained on images containing these specific artifacts will faithfully reproduce them even in clean brain scans (indicated by red arrows). These hallucinations are often "stable," meaning they persist even if the input is slightly perturbed, making them dangerously deceptive for medical professionals.
Conclusion
This work serves as a warning that the pursuit of "pretty" images through deep learning can come at the cost of reliability. To develop robust AI-inspired methods, we must move toward hybrid approaches that combine DL with classical model-based regularization and new training paradigms that explicitly enforce consistency with the physics of the problem ($A\Psi(y) \approx y$).
Summary updated