Abstract
We propose a theory that relates difficulty of learning in deep architectures to culture and language. It is articulated around the following hypotheses: (1) learning in an individual human brain is hampered by the presence of effective local minima; (2) this optimization difficulty is particularly important when it comes to learning higher-level abstractions, i.e., concepts that cover a vast and highly-nonlinear span of sensory configurations; (3) such high-level abstractions are best represented in brains by the composition of many levels of representation, i.e., by deep architectures; (4) a human brain can learn such high-level abstractions if guided by the signals produced by other humans, which act as hints or indirect supervision for these high-level abstractions; and (5), language and the recombination and optimization of mental concepts provide an efficient evolutionary recombination operator, and this gives rise to rapid search in the space of communicable ideas that help humans build up better high-level internal representations of their world. These hypotheses put together imply that human culture and the evolution of ideas have been crucial to counter an optimization difficulty: this optimization difficulty would otherwise make it very difficult for human brains to capture high-level knowledge of the world. The theory is grounded in experimental observations of the difficulties of training deep artificial neural networks. Plausible consequences of this theory for the efficiency of cultural evolution are sketched.
Access this chapter
Tax calculation will be finalised at checkout
Purchases are for personal use only
Notes
- 1.
- 2.
Note that the rewards received by an agent depend on the tasks that it faces, which may be different depending on the biological and social niche that it occupies.
- 3.
Stationary i.i.d case where examples independently come from the same stationary distribution \(P\).
- 4.
In many machine learning algorithms, one minimizes the training error plus a regularization penalty which prevents the learner from simply learning the training examples by heart without good generalization on new examples.
- 5.
Although it is always possible to trivially overfit the top two layers of a deep network by memorizing patterns, this may still happen with very poor training of lower levels, corresponding to poor representation learning.
- 6.
Which ignores the interaction with the other levels, except for receiving input from the level below.
- 7.
Results got worse in terms of generalization error, while training error could be small thanks to capacity in the top few layers.
- 8.
i.e., communicate the concept as a function that associates an indicator of its presence with all compatible sensory configurations.
- 9.
The training criterion is here seen as a function of the learned parameters, as a sum of an error function over a training distribution of examples.
- 10.
“better” in the sense of the survival value they provide, and how well they allow their owner to understand the world around them. Note how this depends on the context (ecological and social niche) and that there may be many good solutions.
- 11.
Remember that a meme is copied in a process of teaching by example which is highly stochastic, due to the randomness in encounters (in which particular percepts serve as examples of the meme) and due to the small number of examples of the meme. This creates a highly variable randomly distorted version of the meme in the learner’s brain.
- 12.
References
D.H. Ackley, G.E. Hinton, T.J. Sejnowski, A learning algorithm for Boltzmann machines. Cogn. Sci. 9, 147–169 (1985)
M.A. Arbib, The Handbook of Brain Theory and Neural Networks (MIT Press, Cambridge, 1995)
Y. Bengio, Learning deep architectures for AI. Found. Trends Mach. Lear. 2(1), 1–127 2009. Also published as a book. Now Publishers, 2009
Y. Bengio, O. Delalleau, On the expressive power of deep architectures, in Proceedings of the 22nd International Conference on Algorithmic Learning Theory, 2011, ed. by J. Kivinen, C. Szepesvári, E. Ukkonen, T. Zeugmann
Y. Bengio, O. Delalleau, C. Simard, Decision trees do not generalize to new variations. Comput. Intell. 26(4), 449–467 (2010)
Y. Bengio, P. Lamblin, D. Popovici, H. Larochelle, Greedy layer-wise training of deep networks, in Advances in Neural Information Processing Systems 19 (NIPS’06), ed. by B. Schölkopf, J. Platt, T. Hoffman (MIT Press, Cambridge, 2007), pp. 153–160
Y. Bengio, Y. LeCun, Scaling learning algorithms towards AI. in Large Scale Kernel Machines, ed. by L. Bottou, O. Chapelle, D. DeCoste, J. Weston (MIT Press, Cambridge, 2007)
Y. Bengio, J. Louradour, R. Collobert, J. Weston, Curriculum learning, in Proceedings of the Twenty-sixth International Conference on Machine Learning (ICML’09), ed. by L. Bottou, M. Littman (ACM, 2009)
L. Bottou, Stochastic learning, in Advanced Lectures on Machine Learning, number LNAI 3176 in Lecture notes in artificial intelligence, ed. by O. Bousquet, U. von Luxburg (Springer, Berlin, 2004), pp. 146–168
M.A. Carreira-Perpiñan, G.E. Hinton, On contrastive divergence learning, in Proceedings of the Tenth International Workshop on Artificial Intelligence and Statistics (AISTATS’05), ed. by R.G. Cowell, Z. Ghahramani (Society for Artificial Intelligence and Statistics, 2005) pp. 33–40.
R. Caruana, Multitask connectionist learning, in Proceedings of the 1993 Connectionist Models Summer School, 1993, pp. 372–379
R. Dawkins, The Selfish Gene (Oxford University Press, London, 1976)
K. Distin, The Selfish Meme (Cambridge University Press, London, 2005)
J.L. Elman, Learning and development in neural networks: the importance of starting small. Cognition 48, 781–799 (1993)
D. Erhan, Y. Bengio, A. Courville, P.-A. Manzagol, P. Vincent, S. Bengio, Why does unsupervised pre-training help deep learning? J. Mach. Lear. Res. 11, 625–660 (2010)
J. Håstad, Almost optimal lower bounds for small depth circuits, in Proceedings of the 18th annual ACM Symposium on Theory of Computing (ACM Press, Berkeley, 1986), pp. 6–20
J. Håstad, M. Goldmann, On the power of small-depth threshold circuits. Comput. Complex. 1, 113–129 (1991)
G.E. Hinton, T.J. Sejnowski, D.H. Ackley, Boltzmann machines: constraint satisfaction networks that learn. Technical Report TR-CMU-CS-84-119, (Dept. of Computer Science, Carnegie-Mellon University, 1984)
G.E. Hinton, Learning distributed representations of concepts, in Proceedings of the Eighth Annual Conference of the Cognitive Science Society (Lawrence Erlbaum, Hillsdale, Amherst 1986, 1986), pp. 1–12
G.E. Hinton, Connectionist learning procedures. Artif. Intell. 40, 185–234 (1989)
G.E. Hinton, S.J. Nowlan, How learning can guide evolution. Complex Syst. 1, 495–502 (1989)
G.E. Hinton, S. Osindero, Y.W. Teh, A fast learning algorithm for deep belief nets. Neural Comput. 18, 1527–1554 (2006)
G.E. Hinton, R. Salakhutdinov, Reducing the dimensionality of data with neural networks. Science 313(5786), 504–507 (2006)
J.H. Holland, Adaptation in Natural and Artificial Systems (University of Michigan Press, Ann Arbor, 1975)
E. Hutchins, B. Hazlehurst, How to invent a lexicon: the development of shared symbols in interaction, in Artificial Societies: The Computer Simulation of Social Life, ed. by N. Gilbert, R. Conte (UCL Press, London, 1995), pp. 157–189
E. Hutchins, B. Hazlehurst, Auto-organization and emergence of shared language structure, in Simulating the Evolution of Language, ed. by A. Cangelosi, D. Parisi (Springer, London, 2002), pp. 279–305
K. Jarrett, K. Kavukcuoglu, M. Ranzato, Y. LeCun, What is the best multi-stage architecture for object recognition? in Proceedings of IEEE International Conference on Computer Vision (ICCV’09), 2009, pp. 2146–2153
F. Khan, X. Zhu, B. Mutlu, How do humans teach: on curriculum learning and teaching dimension, in Advances in Neural Information Processing Systems 24 (NIPS’11), 2011 pp. 1449–1457
K.A. Krueger, P. Dayan, Flexible shaping: how learning in small steps helps. Cognition 110, 380–394 (2009)
H. Larochelle, Y. Bengio, J. Louradour, P. Lamblin, Exploring strategies for training deep neural networks. J. Mach. Lear. Res. 10, 1–40 (2009)
H. Lee, R. Grosse, R. Ranganath, A.Y. Ng, Convolutional deep belief networks for scalable unsupervised learning of hierarchical representations, in Proceedings of the Twenty-sixth International Conference on Machine Learning (ICML’09), ed. by L. Bottou, M. Littman (ACM, Montreal (Qc), Canada, 2009)
J. Martens. Deep learning via Hessian-free optimization, in Proceedings of the Twenty-seventh International Conference on Machine Learning (ICML-10), ed. by L. Bottou, M. Littman (ACM, 2010) pp. 735–742
E. Moritz, Memetic science: I–general introduction. J. Ideas 1, 1–23 (1990)
G.B. Peterson, A day of great illumination: B. F. Skinner’s discovery of shaping. J. Exp. Anal. Behav. 82(3), 317–328 (2004)
R. Raina, A. Battle, H. Lee, B. Packer, A.Y. Ng, Self-taught learning: transfer learning from unlabeled data, in Proceedings of the Twenty-fourth International Conference on Machine Learning (ICML’07), ed. by Z. Ghahramani (ACM, 2007), pp. 759–766
M. Ranzato, C. Poultney, S. Chopra, Y. LeCun, Efficient learning of sparse representations with an energy-based model, in Advances in Neural Information Processing Systems 19 (NIPS’06), ed. by B. Schölkopf, J. Platt, T. Hoffman (MIT Press, 2007) pp. 1137–1144
D.E. Rumelhart, J.L. McClelland, and the PDP Research Group Parallel Distributed Processing Explorations in the Microstructure of Cognition, (MIT Press, Cambridge, 1986)
R. Salakhutdinov, G.E. Hinton, Deep Boltzmann machines, in Proceedings of the Twelfth International Conference on Artificial Intelligence and Statistics (AISTATS 2009), vol. 8, 2009
R. Salakhutdinov, G.E. Hinton, Deep Boltzmann machines. in Proceedings of The Twelfth International Conference on Artificial Intelligence and Statistics (AISTATS’09), vol. 5, 2009, pp. 448–455
R. Salakhutdinov, H. Larochelle, Efficient learning of deep Boltzmann machines, in Proceedings of the Thirteenth International Conference on Artificial Intelligence and Statistics (AISTATS 2010), JMLR W&CP, vol. 9, 2010, pp. 693–700
B.F. Skinner, Reinforcement today. Am. Psychol. 13, 94–99 (1958)
F. Subiaul, J. Cantlon, R.L. Holloway, H.S. Terrace, Cognitive imitation in rhesus macaques. Science 305(5682), 407–410 (2004)
R. Sutton, A. Barto, Reinforcement Learning: An Introduction (MIT Press, Cambridge, 1998)
J. Tenenbaum, V. de Silva, J.C. Langford, A global geometric framework for nonlinear dimensionality reduction. Science 290(5500), 2319–2323 (2000)
L. van der Maaten, G.E. Hinton, Visualizing data using t-sne. J. Mach. Learn. Res. 9, 2579–2605 (2008)
J. Weston, F. Ratle, R. Collobert, Deep learning via semi-supervised embedding, in Proceedings of the Twenty-fifth International Conference on Machine Learning (ICML’08), ed. by W.W. Cohen, A. McCallum, S.T. Roweis (ACM, New York, NY, USA, 2008), pp. 1168–1175
A. Yao, Separating the polynomial-time hierarchy by oracles, in Proceedings of the 26th Annual IEEE Symposium on Foundations of Computer Science, 1985, pp. 1–10
A.L. Yuille, The convergence of contrastive divergences, in Advances in Neural Information Processing Systems 17 (NIPS’04), ed. by L.K. Saul, Y. Weiss, L. Bottou (MIT Press, 2005) pp. 1593–1600
Acknowledgments
The author would like to thank Caglar Gulcehre, Aaron Courville, Myriam Côté, and Olivier Delalleau for useful feedback, as well as NSERC, CIFAR and the Canada Research Chairs for funding.
Author information
Authors and Affiliations
Corresponding author
Editor information
Editors and Affiliations
Rights and permissions
Copyright information
© 2014 Springer-Verlag Berlin Heidelberg
About this chapter
Cite this chapter
Bengio, Y. (2014). Evolving Culture Versus Local Minima. In: Kowaliw, T., Bredeche, N., Doursat, R. (eds) Growing Adaptive Machines. Studies in Computational Intelligence, vol 557. Springer, Berlin, Heidelberg. https://doi.org/10.1007/978-3-642-55337-0_3
Download citation
DOI: https://doi.org/10.1007/978-3-642-55337-0_3
Published:
Publisher Name: Springer, Berlin, Heidelberg
Print ISBN: 978-3-642-55336-3
Online ISBN: 978-3-642-55337-0
eBook Packages: EngineeringEngineering (R0)