Elamur Research⌕
☰
home / papers / 2018 / Deep Learning: A Critical Appraisal

Deep Learning: A Critical Appraisal

Paper · 2018 · Gary Marcus

Summary

This position paper by Gary Marcus examines deep learning about five years after its 2012 revival, which was sparked by the ImageNet convolutional network of Krizhevsky, Sutskever and Hinton. It asks how much the approach can deliver and whether it can lead to artificial general intelligence. The paper presents no new data or model. Instead, Marcus describes deep learning as statistical pattern classification with multi-layer networks trained by back-propagation and gradient descent. He contrasts this with how humans learn.

His central argument is that these systems interpolate well within large, stable datasets but struggle to extrapolate beyond their training distribution. He lists ten concerns, beginning with their heavy need for data. He concludes that deep learning should be kept but supplemented with other methods, such as symbolic representations of variables, innate priors, common-sense knowledge and narrative understanding.

The paper matters because it gave both researchers and non-specialists a reference critique at the peak of enthusiasm. It also proposed harder benchmarks that go beyond supervised perceptual classification.

Key points

  • The 2012 ImageNet model had nine layers, 60 million parameters and 650,000 nodes. It was trained on about one million examples across about 1,000 categories, with data augmentation through shifts, mirroring and brightness changes.
  • Generalization is framed as interpolation versus extrapolation. Humans, including 7-month-old infants, can pick up abstract rules from a handful of unlabeled examples in about two minutes (Marcus et al., 1999). Deep learning has no way to learn from explicit definitions.
  • The proposed challenge tasks are:
    • open-ended comprehension of video, text or podcasts
    • the Allen AI 8th grade science challenge
    • general game playing with transfer between games
    • an embodied robot that builds objects such as IKEA shelves from instructions
  • Suggested research directions include innate structure (drawing on Marcus 2001 and Spelke), common-sense and physical reasoning, and narrative understanding (Schank & Abelson).
  • Similar doubts have been raised by Lake & Baroni, Chollet, Jia & Liang, Vicarious and Berkeley. Hinton himself said in an Axios interview that he is wary of back-propagation's reliance on labeled data.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓New York UniversityNew York · United States201 works · 100,180 citations

 Fields, topics and concepts

📚 Shelves 2The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 works

💡 Concepts 1Key ideas and methods explained on our concept pages.Artificial intelligence · 49 works

Cite

Cite as Gary Marcus. 2018. Deep Learning: A Critical Appraisal. arXiv:1801.00631.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.1801.00631 · arXiv 1801.00631

@misc{marcus2018deep,
  title = {{Deep Learning: A Critical Appraisal}},
  author = {Gary Marcus},
  year = {2018},
  doi = {10.48550/arxiv.1801.00631},
  eprint = {1801.00631},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.1801.00631}
}

References

79 references, 13 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Athalye, A., Engstrom, L., Ilyas, A., & Kwok, K. (2017). Synthesizing Robust Adversarial Examples. arXiv, cs.CV. find
  2. Besold, T. R., Garcez, A. D., Bader, S., Bowman, H., Domingos, P., Hitzler, P. et al. (2017). Neural-Symbolic Learning and Reasoning: A Survey and Interpretation. arXiv, cs.AI. find
  3. Bošnjak, M., Rocktäschel, T., Naradowsky, J., & Riedel, S. (2016). Programming with a Differentiable Forth Interpreter. arXiv. find
  4. Bottou, L. (2015). Two big challenges in machine learning. Proceedings from 32nd International Conference on Machine Learning. find
  5. Bowman, S. R., Angeli, G., Potts, C., & Manning, C. D. (2015). A large annotated corpus for learning natural language inference. arXiv, cs.CL.
  6. Chollet, F. (2017). Deep Learning with Python. Manning Publications. find
  7. Cireşan, D., Meier, U., Masci, J., & Schmidhuber, J. (2012). Multi-column deep neural network for traffic sign classification. Neural networks. find
  8. Davis, E., & Marcus, G. (2015). Commonsense reasoning and commonsense knowledge in artificial intelligence. Communications of the ACM, 58(9)(9), 92-103.
  9. Davis, E. (2016). How to Write Science Questions that Are Easy for People and Hard for Computers. AI magazine, 37(1)(1), 13-22. find
  10. Davis, E., Marcus, G., & Frazier-Logue, N. (2017). Commonsense reasoning about containers using radically incomplete information. Artificial Intelligence, 248, 46-84. find
  11. Deng, J., Dong, W., Socher, R., Li, L. J., Li - Computer Vision and, K., & 2009 Imagenet: A large-scale hierarchical image database. Proceedings from Computer Vision and Pattern Recognition, 2009. find
  12. CVPR 2009. IEEE Conference on. find
  13. Elman, J. L. (1990). Finding structure in time. Cognitive science, 14(2)(2), 179-211. find
  14. Evtimov, I., Eykholt, K., Fernandes, E., Kohno, T., Li, B., Prakash, A. et al. (2017). Robust Physical-World Attacks on Deep Learning Models. arXiv, cs.CR. find
  15. Fodor, J. A., & Pylyshyn, Z. W. (1988). Connectionism and cognitive architecture: a critical analysis. Cognition, 28(1-2)(1-2), 3-71. find
  16. Gardner, H. (2011). Frames of mind: The theory of multiple intelligences. Basic books. find
  17. Gelman, S. A., Leslie, S. J., Was, A. M., & Koch, C. M. (2015). Children’s interpretations of general quantifiers, specific quantifiers, and generics. Lang Cogn Neurosci, 30(4)(4), 448-461. find
  18. Genesereth, M., Love, N., & Pell, B. (2005). General game playing: Overview of the AAAI competition. AI magazine, 26(2)(2), 62. find
  19. George, D., Lehrach, W., Kansky, K., Lázaro-Gredilla, M., Laan, C., Marthi, B. et al. (2017). A generative vision model that trains with high data efficiency and breaks text-based CAPTCHAs. Science, 358(6368)(6368). find
  20. Gervain, J., Berent, I., & Werker, J. F. (2012). Binding at birth: the newborn brain detects identity relations and sequential position in speech. J Cogn Neurosci, 24(3)(3), 564-574. find
  21. Goodfellow, I., Bengio, Y., & Courville, A. (2016). Deep learning. MIT press. find
  22. Graves, A., Wayne, G., Reynolds, M., Harley, T., Danihelka, I., Grabska-Barwińska, A. et al. (2016). Hybrid computing using a neural network with dynamic external memory. Nature, 538(7626)(7626), 471-476. find
  23. Henderson, P., Islam, R., Bachman, P., Pineau, J., Precup, D., & Meger, D. (2017). Deep Reinforcement Learning that Matters. arXiv, cs.LG. find
  24. Huang, S., Papernot, N., Goodfellow, I., Duan, Y., & Abbeel, P. (2017). Adversarial Attacks on Neural Network Policies. arXiv, cs.LG. find
  25. Jia, R., & Liang, P. (2017). Adversarial Examples for Evaluating Reading Comprehension Systems. arXiv.
  26. Kahneman, D. (2013). Thinking, fast and slow (1st pbk. ed. ed.). New York: Farrar, Straus and Giroux. find
  27. Kansky, K., Silver, T., Mély, D. A., Eldawy, M., Lázaro-Gredilla, M., Lou, X. et al. (2017). Schema Networks: Zero-shot Transfer with a Generative Causal Model of Intuitive Physics. arXIv, cs.AI. find
  28. Kočiský, T., Schwarz, J., Blunsom, P., Dyer, C., Hermann, K. M., Melis, G. et al. (2017). The NarrativeQA Reading Comprehension Challenge. arXiv, cs.CL.
  29. Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2012). Imagenet classification with deep convolutional neural networks. In (pp. 1097-1105).
  30. Lake, B. M., Salakhutdinov, R., & Tenenbaum, J. B. (2015). Human-level concept learning through probabilistic program induction. Science, 350(6266)(6266), 1332-1338. find
  31. Lake, B. M., Ullman, T. D., Tenenbaum, J. B., & Gershman, S. J. (2016). Building Machines That Learn and Think Like People. Behav Brain Sci, 1-101. find
  32. Lake, B. M., & Baroni, M. (2017). Still not systematic after all these years: On the compositional skills of sequence-to-sequence recurrent networks. arXiv. find
  33. Lazer, D., Kennedy, R., King, G., & Vespignani, A. (2014). Big data. The parable of Google Flu: traps in big data analysis. Science, 343(6176)(6176), 1203-1205. find
  34. Le, Q. V., Ranzato, M.-A., Monga, R., Devin, M., Chen, K., Corrado, G. et al. (2012). Building high-level features using large scale unsupervised learning. Proceedings from International Conference on Machine Learning. find
  35. LeCun, Y. (1989). Generalization and network design strategies. Technical Report CRG-TR-89-4. find
  36. Lerer, A., Gross, S., & Fergus, R. (2016). Learning Physical Intuition of Block Towers by Example. arXiv, cs.AI. find
  37. Lighthill, J. (1973). Artificial Intelligence: A General Survey. Artificial Intelligence: a paper symposium. find
  38. Lipton, Z. C. (2016). The Mythos of Model Interpretability. arXiv, cs.LG. find
  39. Lopez-Paz, D., Nishihara, R., Chintala, S., Schölkopf, B., & Bottou, L. (2017). Discovering causal signals in images. Proceedings from Proceedings of Computer Vision and Pattern Recognition (CVPR). find
  40. Luc, P., Neverova, N., Couprie, C., Verbeek, J., & LeCun, Y. (2017). Predicting Deeper into the Future of Semantic Segmentation. International Conference on Computer Vision (ICCV 2017). find
  41. Page 25 ! of ! 27 Marcus, G., Rossi, F., Veloso - AI Magazine, M., & 2016. (2016). Beyond the Turing Test. AI Magazine, Whole issue. find
  42. Marcus, G., Marblestone, A., & Dean, T. (2014). The atoms of neural computation. Science, 346(6209)(6209), 551-552. find
  43. Marcus, G. (in prep). Innateness, AlphaZero, and Artificial Intelligence. find
  44. Marcus, G. (2014). What Comes After the Turing Test? The New Yorker. find
  45. Marcus, G. (2012). Is “Deep Learning” a Revolution in Artificial Intelligence? The New Yorker. find
  46. Marcus, G. F. (2008). Kluge : the haphazard construction of the human mind. Boston: Houghton Mifflin. find
  47. Marcus, G. F. G. F. (2001). The Algebraic Mind: Integrating Connectionism and cognitive science. Cambridge, Mass.: MIT Press. find
  48. Marcus, G. F. (1998a). Rethinking eliminative connectionism. Cogn Psychol, 37(3)(3), 243-282. find
  49. Marcus, G. F. (1998b). Can connectionism save constructivism? Cognition, 66(2)(2), 153-182. find
  50. Marcus, G. F., Pinker, S., Ullman, M., Hollander, M., Rosen, T. J., & Xu, F. (1992). Overregularization in language acquisition. Monogr Soc Res Child Dev, 57(4)(4), 1-182. find
  51. Marcus, G. F., Vijayan, S., Bandi Rao, S., & Vishton, P. M. (1999). Rule learning by seven-month-old infants. Science, 283(5398)(5398), 77-80. find
  52. Mikolov, T., Chen, K., Corrado, G., & Dean, J. (2013). Efficient Estimation of Word Representations in Vector Space. arXiv.
  53. Mnih, V., Kavukcuoglu, K., Silver, D., Rusu, A. A., Veness, J., Bellemare, M. G. et al. (2015). Human-level control through deep reinforcement learning. Nature, 518(7540)(7540), 529-533. find
  54. Neelakantan, A., Le, Q. V., Abadi, M., McCallum, A., & Amodei, D. (2016). Learning a Natural Language Interface with Neural Programmer. arXiv. find
  55. Ng, A. (2016). What Artificial Intelligence Can and Can’t Do Right Now. Harvard Business Review. find
  56. Nguyen, A., Clune, J., Bengio, Y., Dosovitskiy, A., & Yosinski, J. (2016). Plug & Play Generative Networks: Conditional Iterative Generation of Images in Latent Space. arXiv, cs.CV. find
  57. Nguyen, A., Yosinski, J., & Clune, J. (2014). Deep Neural Networks are Easily Fooled: High Confidence Predictions for Unrecognizable Images. arXiv, cs.CV. find
  58. Norvig, P. (2016). State-of-the-Art AI: Building Tomorrow’s Intelligent Systems. Proceedings from EmTech Digital, San Francisco. find
  59. O’Neil, C. (2016). Weapons of math destruction : how big data increases inequality and threatens democracy. find
  60. Ortiz Jr, C. L. (2016). Why we need a physically embodied Turing test and what it might look like. AI magazine, 37(1)(1), 55-63. find
  61. Paritosh, P., & Marcus, G. (2016). Toward a comprehension challenge, using crowdsourcing as a tool. AI Magazine, 37(1)(1), 23-31. find
  62. Pearl, J. (2000). Causality : models, reasoning, and inference /. Cambridge, U.K.; New York : Cambridge University Press. find
  63. Pinker, S., & Prince, A. (1988). On language and connectionism: analysis of a parallel distributed processing model of language acquisition. Cognition, 28(1-2)(1-2), 73-193. find
  64. Ribeiro, M. T., Singh, S., & Guestrin, C. (2016). “Why Should I Trust You?”: Explaining the Predictions of Any Classifier. arXiv, cs.LG.
  65. Sabour, S., Frosst, N., & Hinton, G. E. (2017). Dynamic Routing Between Capsules. arXiv, cs.CV. find
  66. Samek, W., Wiegand, T., & Müller, K.-R. (2017). Explainable Artificial Intelligence: Understanding, Visualizing and Interpreting Deep Learning Models. arXiv, cs.AI. find
  67. Schank, R. C., & Abelson, R. P. (1977). Scripts, Plans, Goals and Understanding: an Inquiry into Human Knowledge Structures. Hillsdale, NJ: L. Erlbaum. find
  68. Schmidhuber, J. (2015). Deep learning in neural networks: An overview. Neural networks. find
  69. Schoenick, C., Clark, P., Tafjord, O., P, T., & Etzioni, O. (2017). Moving beyond the Turing Test with the Allen AI Science Challenge. Communications of the ACM, 60 (9)(9), 60-64.
  70. Sculley, D., Phillips, T., Ebner, D., Chaudhary, V., & Young, M. (2014). Machine learning: The high-interest credit card of technical debt. Proceedings from SE4ML: Software Engineering for Machine Learning (NIPS 2014 Workshop). find
  71. Socher, R., Huval, B., Manning, C. D., & Ng, A. Y. (2012). Semantic compositionality through recursive matrix-vector spaces. Proceedings from Proceedings of the 2012 joint conference on empirical methods in natural language processing and computational natural language learning.
  72. Spelke, E. S., & Kinzler, K. D. (2007). Core knowledge. Dev Sci, 10(1)(1), 89-96. find
  73. Stoica, I., Song, D., Popa, R. A., Patterson, D., Mahoney, M. W., Katz, R. et al. (2017). A Berkeley View of Systems Challenges for AI. arXiv, cs.AI. find
  74. Szegedy, C., Zaremba, W., Sutskever, I., Bruna, J., Erhan, D., Goodfellow, I. et al. (2013). Intriguing properties of neural networks. arXiv, cs.CV.
  75. Vinyals, O., Toshev, A., Bengio, S., & Erhan, D. (2014). Show and Tell: A Neural Image Caption Generator. arXiv, cs.CV.
  76. Watters, N., Tacchetti, A., Weber, T., Pascanu, R., Battaglia, P., & Zoran, D. (2017). Visual Interaction Networks. arXiv. find
  77. Williams, A., Nangia, N., & Bowman, S. R. (2017). A Broad-Coverage Challenge Corpus for Sentence Understanding through Inference. arXiv, cs.CL.
  78. Wu, J., Lu, E., Kohli, P., Freeman, B., & Tenenbaum, J. (2017). Learning to See Physics via Visual De-animation. Proceedings from Advances in Neural Information Processing Systems. find
  79. Zoph, B., Vasudevan, V., Shlens, J., & Le, Q. V. (2017). Learning Transferable Architectures for Scalable Image Recognition. arXiv, cs.CV.

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/marcus-2018-deep-learning-critical-appraisal/