Mathematical Algorithm Design for Deep Learning under Societal and Judicial Constraints: The Algorithmic Transparency Requirement (2401.10310v1)
Abstract: Deep learning still has drawbacks in terms of trustworthiness, which describes a comprehensible, fair, safe, and reliable method. To mitigate the potential risk of AI, clear obligations associated to trustworthiness have been proposed via regulatory guidelines, e.g., in the European AI Act. Therefore, a central question is to what extent trustworthy deep learning can be realized. Establishing the described properties constituting trustworthiness requires that the factors influencing an algorithmic computation can be retraced, i.e., the algorithmic implementation is transparent. Motivated by the observation that the current evolution of deep learning models necessitates a change in computing technology, we derive a mathematical framework which enables us to analyze whether a transparent implementation in a computing model is feasible. We exemplarily apply our trustworthiness framework to analyze deep learning approaches for inverse problems in digital and analog computing models represented by Turing and Blum-Shub-Smale Machines, respectively. Based on previous results, we find that Blum-Shub-Smale Machines have the potential to establish trustworthy solvers for inverse problems under fairly general conditions, whereas Turing machines cannot guarantee trustworthiness to the same degree.
- “Improved protein structure prediction using potentials from deep learning” In Nature 577, 2020, pp. 706–710
- “Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification” In ICCV 2015 IEEE, pp. 1026–1034
- “Mastering the game of Go with deep neural networks and tree search” In Nature 529, 2016, pp. 484–503
- “Language Models are Few-Shot Learners” In NeurIPS 2020 33 Curran Associates, Inc., pp. 1877–1901
- “Learning skillful medium-range global weather forecasting” In Science, 2023
- Yann LeCun, Yoshua Bengio and Geoffrey Hinton “Deep learning” In Nature 521, 2015, pp. 436–444
- Ian Goodfellow, Yoshua Bengio and Aaron Courville “Deep Learning” http://www.deeplearningbook.org MIT Press, 2016
- “The Modern Mathematics of Deep Learning” In Mathematical Aspects of Deep Learning Cambridge University Press, 2022
- Warren S. McCulloch and Walter Pitts “A Logical Calculus of the Ideas Immanent in Nervous Activity” In Bull. Math. Biophys. 5.4, 1943, pp. 115–133
- David E. Rumelhart, Geoffrey E. Hinton and Ronald J. Williams “Learning representations by back-propagating errors” In Nature 323, 1986, pp. 533–536
- Alex Krizhevsky, Ilya Sutskever and Geoffrey E. Hinton “ImageNet Classification with Deep Convolutional Neural Networks” In Commun. ACM 60.6 New York, NY, USA: Association for Computing Machinery, 2017, pp. 84–90
- “The transformational role of GPU computing and deep learning in drug discovery” In Nat. Mach. Intell. 4.3, 2022, pp. 211–221
- “A Survey on Deep Learning Hardware Accelerators for Heterogeneous HPC Platforms” In arXiv:2306.15552, 2023
- “TPU v4: An Optically Reconfigurable Supercomputer for Machine Learning with Hardware Support for Embeddings” In ISCA 2023 New York, NY, USA: Association for Computing Machinery
- Anne C. Elster and Tor A. Haugdahl “Nvidia Hopper GPU and Grace CPU Highlights” In Computing in Science & Engineering 24.2, 2022, pp. 95–100
- “Deep Residual Learning for Image Recognition” In CVPR 2016 IEEE, pp. 770–778
- Ashish Vaswani “Attention Is All You Need” In NIPS 2017 30 Curran Associates, Inc.
- “A Cookbook of Self-Supervised Learning” In arXiv:2304.12210, 2023
- “Deep reinforcement learning from human preferences” In NIPS 2017 30 Curran Associates, Inc.
- “Deep Learning’s Diminishing Returns: The Cost of Improvement is Becoming Unsustainable” In IEEE Spectrum 58.10, 2021, pp. 50–55
- “The computational limits of deep learning” In arXiv:2007.05558, 2020
- Semiconductor Research Corporation “The Decadal Plan for Semiconductors” https://www.src.org/about/decadal-plan/, Last accessed 04.01.2024, 2021
- R. Landauer “Irreversibility and Heat Generation in the Computing Process” In IBM J. Res. Dev. 5.3, 1961, pp. 183–191
- “Experimental verification of Landauer’s principle linking information and thermodynamics” In Nature 483, 2012, pp. 187–189
- FACE; Committee on Technology; NSTC “Pioneering the Future Advanced Computing Ecosystem: A Strategic Plan” https://www.nitrd.gov/pubs/Future-Advanced-Computing-Ecosystem-Strategic-Plan-Nov-2020.pdf, Last accessed 04.01.2024, 2020
- “On the Opportunities and Risks of Foundation Models” In pp arXiv:2108.07258, 2022
- “A Survey on Evaluation of Large Language Models” In arXiv:2307.03109, 2023
- Ben Goertzel “Artificial General Intelligence: Concept, State of the Art, and Future Prospects” In J. Artif. Gen. Intell. 5.1, 2014, pp. 1–48
- Max Roser “AI timelines: What do experts in artificial intelligence expect for the future?” https://ourworldindata.org/ai-timelines In Our World in Data, 2023
- “Incorporating Ethical Considerations Into Automated Vehicle Control” In Trans. Intell. Transport. Syst. 18.6 IEEE Press, 2017, pp. 1429–1439
- Stamatis Karnouskos “Self-Driving Car Acceptance and the Role of Ethics” In IEEE Trans. Eng. Manag. 67.2, 2020, pp. 252–265
- “Autonomous Driving Ethics: from Trolley Problem to Ethics of Risk” In Philosophy & Technology 34.4, 2021, pp. 1033–1055
- “Explainable Deep Learning: A Field Guide for the Uninitiated” In J. Artif. Int. Res. 73 El Segundo, CA, USA: AI Access Foundation, 2022
- “Intriguing properties of neural networks” In ICLR 2014
- “Siren’s Song in the AI Ocean: A Survey on Hallucination in Large Language Models” In arXiv:2309.01219, 2023
- Alexander Bastounis, Anders C Hansen and Verner Vlačić “The mathematics of adversarial attacks in AI – Why deep learning is unstable despite the existence of stable neural networks” In arXiv:2109.06098, 2021
- “The Gap between Theory and Practice in Function Approximation with Deep Neural Networks” In SIAM J. Math. Data Sci. 3.2, 2021, pp. 624–655
- “Adversarial Examples Are Not Bugs, They Are Features” In NeurIPS 2019 Red Hook, NY, USA: Curran Associates Inc.
- “Audio Adversarial Examples: Targeted Attacks on Speech-to-Text” In SPW 2018 IEEE, pp. 1–7
- “Robustness May Be at Odds with Accuracy” In ICLR 2019
- “On instabilities of deep learning in image reconstruction and the potential costs of AI” In Proc. Natl. Acad. Sci. 117.48 National Academy of Sciences, 2020, pp. 30088–30095
- Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi and Pascal Frossard “DeepFool: A Simple and Accurate Method to Fool Deep Neural Networks” In CVPR 2016 IEEE, pp. 2574–2582
- Amine Boulemtafes, Abdelouahid Derhab and Yacine Challal “A review of privacy-preserving techniques for deep learning” In Neurocomputing 384, 2020, pp. 21–45
- “Towards Security Threats of Deep Learning Systems: A Survey” In IEEE Trans. Softw. Eng. 48.5, 2022, pp. 1743–1770
- “Privacy and Security Issues in Deep Learning: A Survey” In IEEE Access 9, 2021, pp. 4566–4593
- “Privacy in Deep Learning: A Survey” In arXiv:2004.12254, 2020
- “Safety Concerns and Mitigation Approaches Regarding the Use of Deep Learning in Safety-Critical Perception Tasks” In SAFECOMP 2020 Workshops Cham: Springer International Publishing, pp. 336–350
- “On 6G and Trustworthiness” In Commun. ACM 65.4 New York, NY, USA: Association for Computing Machinery, 2022, pp. 48–49
- European Parliament “Artificial intelligence act” https://www.europarl.europa.eu/RegData/etudes/BRIE/2021/698792/EPRS_BRI(2021)698792_EN.pdf, 2023
- G7 Hiroshima Summit 2023 “G7 Hiroshima Leaders’ Communiqué” https://www.g7hiroshima.go.jp/documents/pdf/Leaders_Communique_01_en.pdf, 2023
- Haocheng Tan “A brief history and technical review of the expert system research” In IOP Conf. Ser.: Mater. Sci. Eng. 242.1 IOP Publishing, 2017
- Holger Boche, Adalbert Fono and Gitta Kutyniok “Limitations of Deep Learning for Inverse Problems on Digital Hardware” In IEEE Trans. Inf. Theory 69.12, 2023, pp. 7887–7908
- Holger Boche, Adalbert Fono and Gitta Kutyniok “Inverse Problems Are Solvable on Real Number Signal Processing Hardware” In arXiv:2204.02066, 2022
- A.M. Turing “On Computable Numbers, with an Application to the Entscheidungs-problem” In Proc. Lond. Math. Soc. s2-42.1, 1936, pp. 230–265
- Lenore Blum, Mike Shub and Steve Smale “On a theory of computation and complexity over the real numbers: NP-completeness, recursive functions and universal machines” In Bull. (New Ser.) Am. Math. Soc. 21.1 American Mathematical Society, 1989, pp. 1–46
- Bernd Ulmann “Analog Computing” Berlin, Boston: De Gruyter Oldenbourg, 2022
- Wilfried Haensch, Tayfun Gokmen and Ruchir Puri “The Next Generation of Deep Learning Hardware: Analog Computing” In Proc. IEEE 107.1, 2019, pp. 108–122
- Ryan Hamerly, Saumil Bandyopadhyay and Dirk Englund “Asymptotically fault-tolerant programmable photonics” In Nat. Commun. 13.1, 2022
- “Approximate analog computing with metatronic circuits” In Commun. Phys. 4.1 Nature Publishing Group UK London, 2021, pp. 196
- “2022 Roadmap on Neuromorphic Computing and Engineering” 022501 In Neuromorph. Comput. Eng. 2.2, 2022
- European Commission “Regulatory , framework , proposal , on artificial , intelligence” https://digital-strategy.ec.europa.eu/policies/regulatory-framework-ai, Last accessed 04.01.2024
- European Commission “European centre for algorithmic transparency” https://algorithmic-transparency.ec.europa.eu/about_en, Last accessed 04.01.2024
- “Complexity and Real Computation” New York: Springer Verlag, 1998
- B.Jack Copeland “The Church-Turing Thesis” In The Stanford Encyclopedia of Philosophy Metaphysics Research Lab, Stanford University, https://plato.stanford.edu/archives/sum2020/entries/church-turing/, 2020
- S.C. Kleene “General recursive functions of natural numbers.” In Mathematische Annalen 112, 1936, pp. 727–742
- A.M. Turing “Computability and lambda-Definability” In J. Symb. Log. 2.4 Cambridge University Press, 1937, pp. 153–163
- “Computability and analysis: the legacy of Alan Turing” In Turing’s Legacy: Developments from Turing’s Ideas in Logic, Lecture Notes in Logic Cambridge University Press, 2014, pp. 1–47
- Marian B. Pour-El and J.Ian Richards “Computability in Analysis and Physics”, Perspectives in Logic Cambridge University Press, 2017
- Klaus Weihrauch “Computable Analysis: An Introduction” Berlin, Heidelberg: Springer-Verlag, 2000
- Matthew J. Colbrook, Vegard Antun and Anders C. Hansen “The difficulty of computing stable and accurate neural networks: On the barriers of deep learning and Smale’s 18th problem” In Proc. Natl. Acad. Sci. 119.12, 2022
- “Pathways to cellular supremacy in biocomputing” In Nat. Commun. 10, 2019
- “Backpropagation for Energy-Efficient Neuromorphic Computing” In NIPS 2015 28 Curran Associates, Inc.
- “Neuromorphic scaling advantages for energy-efficient random walk computations” In Nat. Electron. 5.2, 2022, pp. 102–112
- “A Long Short-Term Memory for AI Applications in Spike-based Neuromorphic Hardware” In Nat. Mach. Intell. 4.5, 2022, pp. 467–479
- “Physics for neuromorphic computing” In Nat. Rev. Phys. 2.9, 2020, pp. 499–510
- “Event-Driven Signal Processing with Neuromorphic Computing Systems” In ICASSP 2020 IEEE, pp. 8534–8538
- “Opportunities for neuromorphic computing algorithms and applications” In Nat. Comput. Sci. 2, 2022, pp. 10–19
- William Aspray “John von Neumann and the Origins of Modern Computing” Cambridge, MA, USA: MIT Press, 1990
- John Backus “Can Programming Be Liberated from the von Neumann Style? A Functional Style and Its Algebra of Programs” In Commun. ACM 21.8 New York, NY, USA: Association for Computing Machinery, 1978, pp. 613–641
- Danijela Efnusheva, Ana Cholakoska and Aristotel Tentov “A survey of different approaches for overcoming the processor-memory bottleneck” In Int. J. Comput. Sci. Inf. Technol. 9.2, 2017, pp. 151–163
- “Temperature sensitivity of analog in-memory computing using phase-change memory” In IEDM 2021 IEEE
- “In-memory hyperdimensional computing” In Nat. Electron. 3.6, 2020, pp. 327–337
- “Memory devices and applications for in-memory computing” In Nat. Nanotechnol. 15.7, 2020, pp. 529–544
- “A neuromorphic systems approach to in-memory computing with non-ideal memristive devices: from mitigation to exploitation” In Faraday Discuss. 213 The Royal Society of Chemistry, 2019, pp. 487–510
- Ádám Papp, Wolfgang Porod and Gyorgy Csaba “Nanoscale neural network using non-linear spin-wave interference” In Nat. Commun. 12, 2021
- Klaus F. Wagenbauer, Christian Sigl and Hendrik Dietz “Gigadalton-scale shape-programmable DNA assemblies” In Nature 552, 2017, pp. 78–83
- “Illuminating dendritic function with computational models” In Nat. Rev. Neurosci. 21, 2020, pp. 303–321
- “Deep physical neural networks trained with backpropagation” In Nature 601, 2022, pp. 549–555
- “Practical quantum advantage in quantum simulationn” In Nature 607 The Royal Society of Chemistry, 2022, pp. 667–676
- “Propagation of errors and quantitative quantum simulation with quantum advantage” In Quantum Sci. Technol. 7.4 IOP Publishing, 2022
- “Memory effects can make the transmission capability of a communication channel uncomputable” In Nat. Commun. 9.1, 2018
- Rafael F. Schaefer, Holger Boche and H.Vincent Poor “Turing Meets Shannon: On the Algorithmic Computability of the Capacities of Secure Communication Systems (Invited Paper)” In SPAWC 2019 IEEE, pp. 1–5
- “On the Algorithmic Solvability of Spectral Factorization and Applications” In IEEE Trans. Inf. Theory 66.7, 2020, pp. 4574–4592
- Holger Boche and Ullrich J. Mönich “Turing Computability of Fourier Transforms of Bandlimited and Discrete Signals” In IEEE Trans. Signal Process. 68, 2020, pp. 532–547
- Holger Boche and Ullrich J. Mönich “On the Solvability of the Peak Value Problem for Bandlimited Signals With Applications” In IEEE Trans. Signal Process. 69, 2021, pp. 103–118
- Marian B. Pour-El and Ning Zhong “The Wave Equation with Computable Initial Data Whose Unique Solution is nowhere computable” In Math. Log. Q. 43.4, 1997, pp. 499–509
- “Turing Meets Circuit Theory: Not Every Continuous-Time LTI System Can be Simulated on a Digital Computer” In IEEE Trans. Circuits Syst. I: Regul. Pap. 67.12, 2020, pp. 5051–5064
- Holger Boche, Rafael F. Schaefer and H. Vincent Poor “Real Number Signal Processing can Detect Denial-of-Service Attacks” In ICASSP 2021 IEEE, 2021, pp. 4765–4769
- “Detectability of Denial-of-Service Attacks on Arbitrarily Varying Classical-Quantum Channels” In ISIT 2021 IEEE, 2021, pp. 912–917
- Holger Boche, Yannik Böck and Christian Deppe “Deciding the Problem of Remote State Estimation via Noisy Communication Channels on Real Number Signal Processing Hardware” In ICC 2022 IEEE, pp. 4510–4515
- Lenore Blum “Computing over the Reals: Where Turing meets Newton” In Not. Am. Math. Soc. 51.9, 2004, pp. 1024–1034
- Jacob D. Bekenstein “Universal upper bound on the entropy-to-energy ratio for bounded systems” In Phys. Rev. D 23 American Physical Society, 1981, pp. 287–298
- Chris Olah “Mechanistic Interpretability, Variables, and the Importance of Interpretable Bases” https://www.transformer-circuits.pub/2022/mech-interp-essay, Last accessed 04.01.2024, 2022
- “Explaining AI Through Mechanistic Interpretability”, 2023 URL: http://philsci-archive.pitt.edu/22747/
- “A survey of human-in-the-loop for machine learning” In Future Gener. Comput. Syst. 135, 2022, pp. 364–381
- “A Safe, Secure, and Predictable Software Architecture for Deep Learning in Safety-Critical Systems” In IEEE Embed. Syst. Lett. 12.3, 2020, pp. 78–82
- “Towards Stable and Efficient Training of Verifiably Robust Neural Networks” In ICLR 2020
- “Robustness Certification with Generative Models” In SIGPLAN PLDI 2021 New York, NY, USA: Association for Computing Machinery, pp. 1141–1154
- I. Daubechies, M. Defrise and C. De Mol “An iterative thresholding algorithm for linear inverse problems with a sparsity constraint” In Commun. Pure Appl. Math. 57.11, 2004, pp. 1413–1457
- Stephen J. Wright, Robert D. Nowak and MÁrio A.T. Figueiredo “Sparse Reconstruction by Separable Approximation” In IEEE Trans. Signal Process. 57.7, 2009, pp. 2479–2493
- “Sparse solutions to linear inverse problems with multiple measurement vectors” In IEEE Trans. Signal Process. 53.7, 2005, pp. 2477–2488
- Ivan Selesnick “Sparse Regularization via Convex Analysis” In IEEE Trans. Signal Process. 65.17, 2017, pp. 4481–4494
- “Decoding by linear programming” In IEEE Trans. Inf. Theory 51.12, 2005, pp. 4203–4215
- E.J. Candes, J. Romberg and T. Tao “Robust uncertainty principles: exact signal reconstruction from highly incomplete frequency information” In IEEE Trans. Inf. Theory 52.2, 2006, pp. 489–509
- Emmanuel J. Candes and Terence Tao “Near-Optimal Signal Recovery From Random Projections: Universal Encoding Strategies?” In IEEE Trans. Inf. Theory 52.12, 2006, pp. 5406–5425
- D.L. Donoho “Compressed sensing” In IEEE Trans. Inf. Theory 52.4, 2006, pp. 1289–1306
- Shihao Ji, Ya Xue and Lawrence Carin “Bayesian Compressive Sensing” In IEEE Trans. Signal Process. 56.6, 2008, pp. 2346–2356
- Marco F. Duarte and Yonina C. Eldar “Structured Compressed Sensing: from Theory to Applications” In IEEE Trans. Signal Process. 59.9, 2011, pp. 4053–4085
- Michael Elad “Optimized Projections for Compressed Sensing” In IEEE Trans. Signal Process. 55.12, 2007, pp. 5695–5702
- Emmanuel J. Candès, Justin K. Romberg and Terence Tao “Stable Signal Recovery from Incomplete and Inaccurate Measurements” In Commun. Pure Appl. Math. 59.8, 2006, pp. 1207–1223
- Scott Shaobing Chen, David L. Donoho and Michael A. Saunders “Atomic Decomposition by Basis Pursuit” In SIAM J. Sci. Comput. 20.1, 1998, pp. 33–61
- J.A. Tropp “Just relax: convex programming methods for identifying sparse signals in noise” In IEEE Trans. Inf. Theory 52.3, 2006, pp. 1030–1051
- A. Belloni, V. Chernozhukov and L. Wang “Square-Root Lasso: Pivotal Recovery of Sparse Signals via Conic Programming” In Biometrika 98.4, 2011, pp. 791–806
- “Image reconstruction by domain-transform manifold learning” In Nature 555, 2018, pp. 487–492
- “Solving inverse problems using data-driven models” In Acta Numerica 28, 2019, pp. 1–174
- “Learning the Invisible: A Hybrid Deep Learning-Shearlet Framework for Limited Angle Computed Tomography” In Inverse Problems 35.6 IOP Publishing, 2019
- “Deep ADMM-Net for Compressive Sensing MRI” In NIPS 2016 29 Curran Associates, Inc.
- “Learning a variational network for reconstruction of accelerated MRI data” In Magn. Reson. Med. 79.6, 2018, pp. 3055–3071
- “Learning to See in the Dark” In CVPR 2018 IEEE
- “Deep learning microscopy” In Optica 4.11 OSA, 2017, pp. 1437–1443
- “Deep-learning tomography” In The Leading Edge 37.1, 2018, pp. 58–66
- “Deep Learning Techniques for Inverse Problems in Imaging” In IEEE J. Sel. Areas Inf. Theory 1.1, 2020, pp. 39–56
- George Cybenko “Approximation by Superpositions of a Sigmoidal Function” In Math. Control Signal 2, 1989, pp. 03–314
- Kurt Hornik “Approximation Capabilities of Multilayer Feedforward Networks” In Neural Networks 4, 1991, pp. 251–257
- “Optimal Approximation with Sparsely Connected Deep Neural Networks” In SIAM J. Math. Data Sci. 1, 2019, pp. 8–45
- “The Power of Depth for Feedforward Neural Networks” In COLT 2016 Proceedings 49, Proceedings of Machine Learning Research PMLR, pp. 907–940
- “Why and when can deep - but not shallow - networks avoid the curse of dimensionality: A review” In Int. J. Autom. Comput. 14, 2017, pp. 503–519
- “Nonlinear Approximation and (Deep) ReLU Networks” In Constr. Approx. 55, 2022, pp. 127–172
- “Approximation spaces of deep neural networks” In Constr. Approx. 55, 2022, pp. 259–367
- Dmitry Yarotsky “Error bounds for approximations with deep ReLU networks” In Neural Networks 94, 2017, pp. 103–114
- “Convolutional Networks for Images, Speech, and Time Series” In The Handbook of Brain Theory and Neural Networks Cambridge, MA, USA: MIT Press, 1998, pp. 255–258
- Xiaolei Lv, Guoan Bi and Chunru Wan “The Group Lasso for Stable Recovery of Block-Sparse Signal Representations” In IEEE Trans. Signal Process. 59.4, 2011, pp. 1371–1382
- Alexander Bastounis, Anders C Hansen and Verner Vlačić “The extended Smale’s 9th problem – On computational barriers and paradoxes in estimation, regularisation, computer-assisted proofs and learning” In arXiv:2110.15734, 2021