Papers
Topics
Authors
Recent
Gemini 2.5 Flash
Gemini 2.5 Flash
184 tokens/sec
GPT-4o
7 tokens/sec
Gemini 2.5 Pro Pro
45 tokens/sec
o3 Pro
4 tokens/sec
GPT-4.1 Pro
38 tokens/sec
DeepSeek R1 via Azure Pro
28 tokens/sec
2000 character limit reached

The troublesome kernel -- On hallucinations, no free lunches and the accuracy-stability trade-off in inverse problems (2001.01258v4)

Published 5 Jan 2020 in cs.LG and cs.CV

Abstract: Methods inspired by AI are starting to fundamentally change computational science and engineering through breakthrough performances on challenging problems. However, reliability and trustworthiness of such techniques is 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 provides a theoretical foundation for these phenomena. We give mathematical explanations for how and when such effects arise in arbitrary reconstruction methods, with several of our results taking 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. Our results trace these effects to the kernel of the forward operator whenever it is nontrivial, but also apply to the case when the forward operator is ill-conditioned. Based on these insights, our work aims to spur research into new ways to develop robust and reliable AI-based methods for inverse problems in imaging.

Definition Search Book Streamline Icon: https://streamlinehq.com
References (82)
  1. Deep neural networks are effective at learning high-dimensional Hilbert-valued functions from limited data. In Proceedings of The Second Annual Conference on Mathematical and Scientific Machine Learning, volume 145 of Proc. Mach. Learn. Res., pages 1–36, 2021.
  2. B. Adcock and N. Dexter. The gap between theory and practice in function approximation with deep neural networks. SIAM J. Math. Data Sci., 3(2):624–655, 2021.
  3. B. Adcock and A. C. Hansen. Compressive Imaging: Structure, Sampling, Learning. Cambridge University Press, Cambridge, UK, 2021.
  4. Localized adversarial artifacts for compressed sensing MRI. arXiv:2206.05289, 2022.
  5. Radiology data from the cancer genome atlas lung adenocarcinoma [TCGA-LUAD] collection. The Cancer Imaging Archive, 10:K9, 2016.
  6. Linear and graphical models: for the multivariate complex normal distribution, volume 101. Springer-Verlag, 1995.
  7. Am (A)I hallucinating? Non-robustness, hallucinations and unpredictable performance of AI for MR image reconstruction. Preprint, 2023.
  8. On instabilities of deep learning in image reconstruction and the potential costs of AI. Proc. Natl. Acad. Sci. USA, 117(48):30088–30095, 2020.
  9. Solving inverse problems using data-driven models. Acta Numer., 28:1–174, 2019.
  10. Quantifying model uncertainty in inverse problems via bayesian deep gradient descent. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 1392–1399. IEEE, 2021.
  11. The mathematics of adversarial attacks in AI – Why deep learning is unstable despite the existence of stable neural networks. arXiv:2109.06098, 2021.
  12. C. Belthangady and L. A. Royer. Applications, promises, and pitfalls of deep learning for fluorescence image reconstruction. Nature methods, 16(12):1215–1225, 2019.
  13. On hallucinations in tomographic image reconstruction. IEEE Trans. Med. Imaging, 40(11):3249–3260, 2021.
  14. Model reduction and neural networks for parametric PDEs. J. Comput. Math., 7:121–157, 2021.
  15. Optimal learning. arXiv:2203.15994, 2022.
  16. Complexity and Real Computation. Springer-Verlag New York, Inc., Secaucus, NJ, USA, 1998.
  17. Compressed sensing using generative models. In International Conference on Machine Learning, Proc. Mach. Learn. Res., pages 537–546. PMLR, 2017.
  18. Fundamental performance limits for ideal decoders in high-dimensional linear inverse problems. IEEE Trans. Inform. Theory, 60(12):7928–7946, 2014.
  19. N. Carlini and D. Wagner. Audio adversarial examples: Targeted attacks on speech-to-text. In 2018 IEEE Security and Privacy Worksh., pages 1–7, 2018.
  20. The cancer imaging archive (TCIA): maintaining and operating a public information repository. Journal of digital imaging, 26:1045–1057, 2013.
  21. Compressed sensing and best k𝑘kitalic_k-term approximation. J. Amer. Math. Soc., 22(1):211–231, 2009.
  22. The difficulty of computing stable and accurate neural networks: On the barriers of deep learning and smale’s 18th problem. Proc. Natl. Acad. Sci. USA, 119(12):e2107151119, 2022.
  23. Measuring robustness in deep learning based compressive sensing. In International Conference on Machine Learning, Proc. Mach. Learn. Res., pages 2433–2444. PMLR, 2021.
  24. SKM-TEA: a dataset for accelerated MRI reconstruction with dense image labels for quantitative clinical evaluation. arXiv:2203.06823, 2022.
  25. European Commission. Europe fit for the digital age: Commission proposes new rules and actions for excellence and trust in artificial intelligence. Press release: https://ec.europa.eu/commission/presscorner/detail/en/IP_21_1682, 21. April 2021. Accessed: 2022-08-09.
  26. Robust physical-world attacks on deep learning visual classification. In IEEE Conference on Computer Vision and Pattern Recognition, pages 1625–1634, 2018.
  27. MGH–USC human connectome project datasets with ultra-high b-value diffusion MRI. Neuroimage, 124:1108–1114, 2016.
  28. Adversarial attacks on medical machine learning. Science, 363(6433):1287–1289, 2019.
  29. GE Healthcare. A bold new world: GE healthcare’s pioneering deep learning image reconstruction technology benefits more than 2 million patients globally. Press release: https://www.ge.com/news/press-releases/a-bold-new-world-ge-healthcares-pioneering-deep-learning-image-reconstruction, 6. May 2022. Accessed: 2022-09-22.
  30. Numerical solution of the parametric diffusion equation by deep neural networks. J. Sci. Comput., 88:22, 2021.
  31. Near-exact recovery for tomographic inverse problems via deep learning. In K. Chaudhuri, S. Jegelka, L. Song, C. Szepesvari, G. Niu, and S. Sabato, editors, Proceedings of the 39th International Conference on Machine Learning, volume 162 of Proc. Mach. Learn. Res., pages 7368–7381. PMLR, 17–23 Jul 2022.
  32. Solving inverse problems with deep neural networks – robustness included? IEEE Trans. Pattern Anal. Machine Intellig., 2022.
  33. Model adaptation for inverse problems in imaging. IEEE Trans. Comput. Imaging, 7:661–674, 2021.
  34. A. Gossard and P. Weiss. Training adaptive reconstruction networks for inverse problems. arXiv:2202.11342, 2022.
  35. Revisiting ℓ1subscriptℓ1\ell_{1}roman_ℓ start_POSTSUBSCRIPT 1 end_POSTSUBSCRIPT-wavelet compressed-sensing MRI in the era of deep learning. Proc. Natl. Acad. Sci., 119(33):e2201062119, 2022.
  36. D. Heaven et al. Why deep-learning AIs are so easy to fool. Nature, 574(7777):163–166, 2019.
  37. The promise and peril of deep learning in microscopy. Nature Methods, 18(2):131–132, 2021.
  38. Some investigations on robustness of deep learning in limited angle tomography. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 145–153. Springer, 2018.
  39. Robust compressed sensing using generative models. In Advances in Neural Information Processing Systems, volume 33, pages 713–727. Curran Ass., Inc., 2020.
  40. Deep convolutional neural network for inverse problems in imaging. IEEE Trans. Image Process., 26(9):4509–4522, 2017.
  41. Evaluation of the robustness of learned MR image reconstruction to systematic deviations between training and test data for the models from the fastMRI challenge. In International Workshop on Machine Learning for Medical Image Reconstruction, pages 25–34. Springer, 2021.
  42. Advancing machine learning for MR image reconstruction with an open competition: Overview of the 2019 fastMRI challenge. Magnetic resonance in medicine, (84(6)):3054–3070, 2020.
  43. Avoiding a replication crisis in deep-learning-based bioimage analysis. Nature methods, 18(10):1136–1144, 2021.
  44. Deep magnetic resonance image reconstruction: Inverse problems meet neural networks. IEEE Signal Process. Mag., 37(1):141–151, 2020.
  45. The medical algorithmic audit. The Lancet Digital Health, 2022.
  46. Deep generative adversarial neural networks for compressive sensing MRI. IEEE Trans. Med. Imaging, 38(1):167–179, 2019.
  47. Convolutional neural networks for inverse problems in imaging: A review. IEEE Signal Process. Mag., 34(6):85–95, 11 2017.
  48. M. T. McCann and M. Unser. Biomedical image reconstruction: From the foundations to deep neural networks. Found. Trends Signal Process., 13(3):283–359, 2019.
  49. A survey of optimal recovery, volume Optimal Estimation in Approximation Theory of The IBM Research Symposia Series. Springer, 1977.
  50. Algorithm unrolling: interpretable, efficient deep learning for signal and image processing. IEEE Signal Process. Mag., 38(2):18–44, 2021.
  51. Adversarial robustness of MR image reconstruction under realistic perturbations. In International Workshop on Machine Learning for Medical Image Reconstruction, pages 24–33. Springer, 2022.
  52. Results of the 2020 fastMRI challenge for machine learning MR image reconstruction. IEEE Trans. Med. Imaging, 2021.
  53. S. Mukherjee. A.I. versus M.D. – What happens when diagnosis is automated? The New Yorker, 4 2017.
  54. M. Neyra-Nesterenko and B. Adcock. NESTANets: stable, accurate and efficient neural networks for analysis-sparse inverse problems. Sampl. Theory Signal Process. Data Anal., 21:4, 2023.
  55. Detecting failure modes in image reconstructions with interval neural network uncertainty. International Journal of Computer Assisted Radiology and Surgery, 16(12):2089–2097, 2021.
  56. Deep learning techniques for inverse problems in imaging. IEEE J. Sel. Area. Inform. Theory, 1(1):39–56, 2020.
  57. Optimism in the face of adversity: Understanding and improving deep learning through adversarial robustness. Proc. IEEE, 109(5):635–659, 2021.
  58. Philips. Philips MR SmartSpeed: Increased speed and image quality: Driven by speed and artificial intelligence. Webpage: https://www.usa.philips.com/healthcare/resources/landing/smartspeed, 2022. Accessed: 2022-09-22.
  59. Evaluation and development of deep neural networks for image super-resolution in optical microscopy. Nature Methods, 18(2):194–202, 2021.
  60. Physics-informed neural networks: a deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. J. Comput. Phys., 378:686–707, 2019.
  61. Improving robustness of deep-learning-based image reconstruction. In H. D. III and A. Singh, editors, Proceedings of the 37th International Conference on Machine Learning, volume 119 of Proc. Mach. Learn. Res., pages 7932–7942. PMLR, 13–18 Jul 2020.
  62. XPDNet for MRI reconstruction: An application to the 2020 fastMRI challenge. In 2021 ISMRM annual meeting, volume 275, 2021.
  63. Image reconstruction: From sparsity to data-adaptive methods and machine learning. Proc. IEEE, 108(1):86–109, 2020.
  64. Deep learning for PET image reconstruction. IEEE Trans. Radi. Plasma Med. Sci., 5(1):1–25, 2021.
  65. Data-driven discovery of partial differential equations. Sci. Adv., 3(4), 2017.
  66. A deep cascade of convolutional neural networks for MR image reconstruction. In Int. Conf. Inf. Proc. Med. Imaging, pages 647–658. Springer, 2017.
  67. S. Shalev-Shwartz and S. Ben-David. Understanding Machine Learning - From Theory to Algorithms. Cambridge University Press, 2014.
  68. NeRP: Implicit neural representation learning with prior embedding for sparsely sampled image reconstruction. IEEE Trans. Neur. Net. Lear., pages 1–13, 2022.
  69. Implicit data crimes: Machine learning bias arising from misuse of public data. Proc. Natl. Acad. Sci., 119(13):e2117203119, 2022.
  70. End-to-end variational networks for accelerated MRI reconstruction. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 64–73. Springer, 2020.
  71. R. Strack. Imaging: AI transforms image reconstruction. Nature Methods, 15(5):309, 2018.
  72. Intriguing properties of neural networks. In Proceedings of the International Conference on Learning Representations, 2014.
  73. A mean-field variational inference approach to deep image prior for inverse problems in medical imaging. In Proceedings of the Fourth Conference on Medical Imaging with Deep Learning, volume 143 of Proc. Mach. Learn. Res., pages 745–760. PMLR, 07–09 Jul 2021.
  74. Deep image prior. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2018.
  75. G. Varoquaux and V. Cheplygina. Machine learning for medical imaging: methodological failures and recommendations for the future. NPJ digital medicine, 5(1):1–8, 2022.
  76. Deep learning for tomographic image reconstruction. Nature Machine Intelligence, 2(12):737–748, 2020.
  77. Image reconstruction is a new frontier of machine learning. IEEE Trans. Med. Imaging, 37(6):1289–1296, 2018.
  78. How medical AI devices are evaluated: limitations and recommendations from an analysis of FDA approvals. Nature Medicine, 27(4):582–584, 2021.
  79. Validation and generalizability of self-supervised image reconstruction methods for undersampled MRI. arXiv:2201.12535, 2022.
  80. Instabilities in conventional multi-coil MRI reconstruction with small adversarial perturbations. In 2021 55th Asilomar Conference on Signals, Systems, and Computers. IEEE, 2021.
  81. fastMRI+, clinical pathology annotations for knee and brain fully sampled magnetic resonance imaging data. Scientific Data, 9(1), 2022.
  82. Image reconstruction by domain-transform manifold learning. Nature, 555(7697):487, 03 2018.
Citations (26)
View on Semantic Scholar

Summary

We haven't generated a summary for this paper yet.

Ai Sparkles Streamline Icon: https://streamlinehq.com Summarize Now