Skip to main content

Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.

  • Article
  • Published:

Hybrid computing using a neural network with dynamic external memory

Abstract

Artificial neural networks are remarkably adept at sensory processing, sequence learning and reinforcement learning, but are limited in their ability to represent variables and data structures and to store data over long timescales, owing to the lack of an external memory. Here we introduce a machine learning model called a differentiable neural computer (DNC), which consists of a neural network that can read from and write to an external memory matrix, analogous to the random-access memory in a conventional computer. Like a conventional computer, it can use its memory to represent and manipulate complex data structures, but, like a neural network, it can learn to do so from data. When trained with supervised learning, we demonstrate that a DNC can successfully answer synthetic questions designed to emulate reasoning and inference problems in natural language. We show that it can learn tasks such as finding the shortest path between specified points and inferring the missing links in randomly generated graphs, and then generalize these tasks to specific graphs such as transport networks and family trees. When trained with reinforcement learning, a DNC can complete a moving blocks puzzle in which changing goals are specified by sequences of symbols. Taken together, our results demonstrate that DNCs have the capacity to solve complex, structured tasks that are inaccessible to neural networks without external read–write memory.

This is a preview of subscription content, access via your institution

Access options

Buy this article

Prices may be subject to local taxes which are calculated during checkout

Figure 1: DNC architecture.
Figure 2: Graph tasks.
Figure 3: Traversal on the London Underground.
Figure 4: Mini-SHRDLU analysis.
Figure 5: Mini-SHRDLU results.

Similar content being viewed by others

References

  1. Krizhevsky, A., Sutskever, I. & Hinton, G. E. Imagenet classification with deep convolutional neural networks. In Advances in Neural Information Processing Systems Vol. 25 (eds Pereira, F. et al.) 1097–1105 (Curran Associates, 2012)

    Google Scholar 

  2. Graves, A. Generating sequences with recurrent neural networks. Preprint at http://arxiv.org/abs/1308.0850 (2013)

  3. Sutskever, I., Vinyals, O. & Le, Q. V. Sequence to sequence learning with neural networks. In Advances in Neural Information Processing Systems Vol. 27 (eds Ghahramani, Z. et al.) 3104–3112 (Curran Associates, 2014)

    Google Scholar 

  4. Mnih, V. et al. Human-level control through deep reinforcement learning. Nature 518, 529–533 (2015)

    Article  ADS  CAS  Google Scholar 

  5. Gallistel, C. R. & King, A. P. Memory and the Computational Brain: Why Cognitive Science Will Transform Neuroscience (John Wiley & Sons, 2011)

  6. Marcus, G. F. The Algebraic Mind: Integrating Connectionism and Cognitive Science (MIT Press, 2001)

  7. Kriete, T., Noelle, D. C., Cohen, J. D. & O’Reilly, R. C. Indirection and symbol-like processing in the prefrontal cortex and basal ganglia. Proc. Natl Acad. Sci. USA 110, 16390–16395 (2013)

    Article  ADS  CAS  Google Scholar 

  8. Hinton, G. E. Learning distributed representations of concepts. In Proc. Eighth Annual Conference of the Cognitive Science Society Vol. 1, 1–12 (Lawrence Erlbaum Associates, 1986)

    Google Scholar 

  9. Bottou, L. From machine learning to machine reasoning. Mach. Learn. 94, 133–149 (2014)

    Article  MathSciNet  Google Scholar 

  10. Fusi, S., Drew, P. J. & Abbott, L. F. Cascade models of synaptically stored memories. Neuron 45, 599–611 (2005)

    Article  CAS  Google Scholar 

  11. Ganguli, S., Huh, D. & Sompolinsky, H. Memory traces in dynamical systems. Proc. Natl Acad. Sci. USA 105, 18970–18975 (2008)

    Article  ADS  CAS  Google Scholar 

  12. Kanerva, P. Sparse Distributed Memory (MIT press, 1988)

  13. Amari, S.-i . Characteristics of sparsely encoded associative memory. Neural Netw. 2, 451–457 (1989)

    Article  Google Scholar 

  14. Weston, J., Chopra, S. & Bordes, A. Memory networks. Preprint at http://arxiv.org/abs/1410.3916 (2014)

  15. Vinyals, O., Fortunato, M. & Jaitly, N. Pointer networks. In Advances in Neural Information Processing Systems Vol. 28 (eds Cortes, C et al.) 2692–2700 (Curran Associates, 2015)

    Google Scholar 

  16. Graves, A., Wayne, G. & Danihelka, I. Neural Turing machines. Preprint at http://arxiv.org/abs/1410.5401 (2014)

  17. Bahdanau, D., Cho, K. & Bengio, Y. Neural machine translation by jointly learning to align and translate. Preprint at http://arxiv.org/abs/1409.0473 (2014)

  18. Gregor, K., Danihelka, I., Graves, A., Rezende, D. J. & Wierstra, D. DRAW: a recurrent neural network for image generation. In Proc. 32nd International Conference on Machine Learning (eds Bach, F. & Blei, D. ) 1462 –1471 (JMLR, 2015)

  19. Hintzman, D. L. MINERVA 2: a simulation model of human memory. Behav. Res. Methods Instrum. Comput. 16, 96–101 (1984)

    Article  Google Scholar 

  20. Kumar, A. et al. Ask me anything: dynamic memory networks for natural language processing. Preprint at http://arxiv.org/abs/1506.07285 (2015)

  21. Sukhbaatar, S. et al. End-to-end memory networks. In Advances in Neural Information Processing Systems Vol. 28 (eds Cortes, C et al.) 2431–2439 (Curran Associates, 2015)

    Google Scholar 

  22. Magee, J. C. & Johnston, D. A synaptically controlled, associative signal for Hebbian plasticity in hippocampal neurons. Science 275, 209–213 (1997)

    Article  CAS  Google Scholar 

  23. Johnston, S. T., Shtrahman, M., Parylak, S., GonÇalves, J. T. & Gage, F. H. Paradox of pattern separation and adult neurogenesis: a dual role for new neurons balancing memory resolution and robustness. Neurobiol. Learn. Mem. 129, 60–68 (2016)

    Article  Google Scholar 

  24. O’Reilly, R. C. & McClelland, J. L. Hippocampal conjunctive encoding, storage, and recall: avoiding a trade-off. Hippocampus 4, 661–682 (1994)

    Article  Google Scholar 

  25. Howard, M. W. & Kahana, M. J. A distributed representation of temporal context. J. Math. Psychol. 46, 269–299 (2002)

    Article  MathSciNet  Google Scholar 

  26. Weston, J., Bordes, A., Chopra, S. & Mikolov, T. Towards AI-complete question answering: a set of prerequisite toy tasks. Preprint at http://arxiv.org/abs/1502.05698 (2015)

  27. Hochreiter, S. & Schmidhuber, J. Long short-term memory. Neural Comput. 9, 1735–1780 (1997)

    Article  CAS  Google Scholar 

  28. Bengio, Y., Louradour, J., Collobert, R. & Weston, J. Curriculum learning. In Proc. 26th International Conference on Machine Learning (eds Bottou, L. & Littman, M. ) 41–48 (ACM, 2009)

  29. Zaremba, W. & Sutskever, I. Learning to execute. Preprint at http://arxiv.org/abs/1410.4615 (2014)

  30. Winograd, T. Procedures as a Representation for Data in a Computer Program for Understanding Natural Language. Report No. MAC-TR-84 (DTIC, MIT Project MAC, 1971)

  31. Epstein, R., Lanza, R. P. & Skinner, B. F. Symbolic communication between two pigeons (Columba livia domestica). Science 207, 543–545 (1980)

    Article  ADS  CAS  Google Scholar 

  32. McClelland, J. L., McNaughton, B. L. & O’Reilly, R. C. Why there are complementary learning systems in the hippocampus and neocortex: insights from the successes and failures of connectionist models of learning and memory. Psychol. Rev. 102, 419–457 (1995)

    Article  Google Scholar 

  33. Kumaran, D., Hassabis, D. & McClelland, J. L. What learning systems do intelligent agents need? Complementary learning systems theory updated. Trends Cogn. Sci. 20, 512–534 (2016)

    Article  Google Scholar 

  34. McClelland, J. L. & Goddard, N. H. Considerations arising from a complementary learning systems perspective on hippocampus and neocortex. Hippocampus 6, 654–665 (1996)

    Article  CAS  Google Scholar 

  35. Lake, B. M., Salakhutdinov, R. & Tenenbaum, J. B. Human-level concept learning through probabilistic program induction. Science 350, 1332–1338 (2015)

    Article  ADS  MathSciNet  CAS  Google Scholar 

  36. Rezende, D. J., Mohamed, S., Danihelka, I., Gregor, K. & Wierstra, D. One-shot generalization in deep generative models. In Proc. 33nd International Conference on Machine Learning (eds Balcan, M. F. & Weinberger, K. Q. ) 1521–1529 (JMLR, 2016)

  37. Santoro, A., Bartunov, S., Botvinick, M., Wierstra, D. & Lillicrap, T. Meta-learning with memory-augmented neural networks. In Proc. 33nd International Conference on Machine Learning (eds Balcan, M. F. & Weinberger, K. Q. ) 1842–1850 (JMLR, 2016)

  38. Oliva, A. & Torralba, A. The role of context in object recognition. Trends Cogn. Sci. 11, 520–527 (2007)

    Article  Google Scholar 

  39. Hermann, K. M. et al. Teaching machines to read and comprehend. In Advances in Neural Information Processing Systems Vol. 28 (eds Cortes, C. et al.) 1693–1701 (Curran Associates, 2015)

    Google Scholar 

  40. O’Keefe, J. & Nadel, L. The Hippocampus as a Cognitive Map (Oxford Univ. Press, 1978)

  41. Graves, A., Mohamed, A.-r. & Hinton, G. Speech recognition with deep recurrent neural networks. In IEEE International Conference on Acoustics, Speech and Signal Processing (eds Ward, R. et al.) 6645–6649 (Curran Associates, 2013)

  42. Wilson, P. R., Johnstone, M. S., Neely, M. & Boles, D. Dynamic storage allocation: a survey and critical review. In Memory Management (ed. Baler, H. G. ) 1–116 (Springer, 1995)

  43. Ross, S., Gordon, G. J. & Bagnell, J. A. A reduction of imitation learning and structured prediction to no-regret online learning. In Proc. Fourteenth International Conference on Artificial Intelligence and Statistics (eds Gordon, G. et al.) 627–635 (JMLR, 2010)

  44. Daumé, H. III, Langford, J. & Marcu, D. Search-based structured prediction. Mach. Learn. 75, 297–325 (2009)

    Article  Google Scholar 

  45. Williams, R. J. Simple statistical gradient-following algorithms for connectionist reinforcement learning. Mach. Learn. 8, 229–256 (1992)

    MATH  Google Scholar 

  46. Sutton, R. S., McAllester, D., Singh, S. P. & Mansour, Y. Policy gradient methods for reinforcement learning with function approximation. In Advances in Neural Information Processing Systems Vol. 12 (eds Solla, S. A. et al.) 1057–1063 (MIT Press, 1999)

    Google Scholar 

  47. Schulman, J., Moritz, P., Levine, S., Jordan, M. & Abbeel, P. High-dimensional continuous control using generalized advantage estimation. Preprint at http://arxiv.org/abs/1506.02438 (2015)

  48. van der Maaten, L. & Hinton, G. Visualizing data using t-SNE. J. Mach. Learn. Res. 9, 2579–2605 (2008)

    MATH  Google Scholar 

  49. Dean, J. et al. Large scale distributed deep networks. In Advances in Neural Information Processing Systems Vol. 25 (eds Pereira, F. et al.) 1223–1231 (Curran Associates, 2012)

    Google Scholar 

  50. Werbos, P. J. Backpropagation through time: what it does and how to do it. Proc. IEEE 78, 1550–1560 (1990)

    Article  Google Scholar 

  51. Tieleman, T. & Hinton, G. RmsProp: divide the gradient by a running average of its recent magnitude. Lecture 6.5 of Neural Networks for Machine Learning (COURSERA, 2012); available at http://www.cs.toronto.edu/~tijmen/csc321/slides/lecture_slides_lec6.pdf

Download references

Acknowledgements

We thank D. Silver, M. Botvinick and S. Legg for reviewing the paper prior to submission; P. Dayan, D. Wierstra, G. Hinton, J. Dean, N. Kalchbrenner, J. Veness, I. Sutskever, V. Mnih, A. Mnih, D. Kumaran, N. de Freitas, L. Sifre, R. Pascanu, T. Lillicrap, J. Rae, A. Senior, M. Denil, T. Kocisky, A. Fidjeland, K. Gregor, A. Lerchner, C. Fernando, D. Rezende, C. Blundell and N. Heess for discussions; J. Besley for legal assistance; the rest of the DeepMind team for support and encouragement; and Transport for London for allowing us to reproduce portions of the London Underground map.

Author information

Authors and Affiliations

Authors

Contributions

A.G. and G.W. conceived the project. A.G., G.W., M.R., T.H., I.D., S.G. and E.G. implemented networks and tasks. A.G., G.W., M.R., T.H., A.G.-B., T.R. and J.A. performed analysis. M.R., T.H., I.D., E.G., K.M.H., C.S., P.B., K.K. and D.H. contributed ideas. A.C. prepared graphics. A.G., G.W., M.R., T.H., S.G., A.P.B., Y.Z., G.O. and K.K. performed experiments. A.G., G.W., H.K., K.K. and D.H. managed the project. A.G., G.W., M.R., T.H., K.K. and D.H. wrote the paper.

Corresponding authors

Correspondence to Alex Graves or Greg Wayne.

Ethics declarations

Competing interests

The authors declare no competing financial interests.

Additional information

Reviewer Information Nature thanks Y. Bengio, J. McClelland and the other anonymous reviewer(s) for their contribution to the peer review of this work.

Extended data figures and tables

Extended Data Figure 1 Dynamic memory allocation.

We trained the DNC on a copy problem, in which a series of 10 random sequences was presented as input. After each input sequence was presented, it was recreated as output. Once the output was generated, that input sequence was not needed again and could be erased from memory. We used a DNC with a feedforward controller and a memory of 10 locations—insufficient to store all 50 input vectors with no overwriting. The goal was to test whether the memory allocation system would be used to free and re-use locations as needed. As shown by the read and write weightings, the same locations are repeatedly used. The free gate is active during the read phases, meaning that locations are deallocated immediately after they are read from. The allocation gate is active during the write phases, allowing the deallocated locations to be re-used.

Extended Data Figure 2 Altering the memory size of a trained network.

A DNC trained on the traversal task with 256 memory locations was tested while varying the number of memory locations and graph triples. The heat map shows the fraction of traversals of length 1–10 performed perfectly by the network, out of a batch of 100. There is a clear correspondence between the number of triples in the graph and the number of memory locations required to solve the task, reflecting our earlier analysis (Fig. 3) that suggests that DNC writes each triple to a separate location in memory. The network appears to exploit all available memory, regardless of how much memory it was trained with. This supports our claim that memory is independent of processing in a DNC, and points to large-scale applications such as knowledge graph processing.

Extended Data Figure 3 Probability of achieving optimal solution.

a, DNC. With 10 goals, the performance of a DNC network with respect to satisfying constraints in minimal time as the minimum number of moves to a goal and the number of constraints in a goal are varied. Performance was highest with a large number of constraints in each goal. b, The performance of an LSTM on the same test.

Extended Data Figure 4 Effect of link matrix sparsity on performance.

We trained the DNC on a copy problem, for which a sequence of length 1–100 of size-6 random binary vectors was given as input, and an identical sequence was then required as output. A feedforward controller was used to ensure that the sequences could not be stored in the controller state. The faint lines show error curves for 20 randomly initialized runs with identical hyper-parameters, with link matrix sparsity switched off (pink), sparsity used with K = 5 (green) and with the link matrix disabled altogether (blue). The bold lines show the mean curve for each setting. The error rate is the fraction of sequences copied with no mistakes out of a batch of 100. There does not appear to be any systematic difference between no sparsity and K = 5. We observed similar behaviour for values of K between 2 and 20 (plots omitted for clarity). The task cannot easily be solved without the link matrix because the input sequence has to be recovered in the correct order. Note the abrupt drops in error for the networks with link matrices: these are the points at which the system learns a copy algorithm that generalizes to longer sequences.

Extended Data Table 1 bAbI best and mean results
Extended Data Table 2 Hyper-parameter settings for bAbI, graph tasks and Mini-SHRDLU
Extended Data Table 3 Curriculum results for graph traversal
Extended Data Table 4 Curriculum results for inference
Extended Data Table 5 Curriculum results for shortest-path task
Extended Data Table 6 Curriculum for Mini-SHRDLU

Related audio

Supplementary information

Supplementary Information

This file contains a glossary of symbols and the complete equations. (PDF 85 kb)

Shortest Path Visualisation

This video shows a DNC successfully finding the shortest path between two nodes in a randomly generated graph. By decoding the memory usage of the DNC (as in Fig. 3) we were able to determine which edges were stored in the memory locations it was reading from and writing to at each timestep. The edges being read are shown in pink on the left, while the edges being written are shown in green on the right; the colour saturation indicates the relative strength of the operation. During the initial query phase, the DNC receives the labels for the start and end goal ("390" and "040" respectively). During the ten step planning phase it attempts to determine the shortest path. During this time it repeatedly reads edges close to or along the path, which are indicated by the grey shaded nodes. Beginning with edges attached to the start and end node, it appears to move further afield as the phase progresses. At the same time it writes to several of the edge locations, perhaps marking those edges as visited. Finally, during the answer phase, it successively reads the outgoing edges from the nodes along the shortest path, allowing it to correctly answer the query. (MOV 2604 kb)

Mini-SHRDLU Visualisation

This video shows a DNC successfully performing a reasoning problem in a blocks world. A sequence of letter-labeled goals (S, K, R, Q, E) is presented to the network one step at a time. Each goal consists of a sequence of defining constraints, presented one constraint per time-step. For example, S is: 6 below 2 (6b2); 2 right of 5 (2r5); 6 right of 1 (6r1); 5 above 1 (5a1). On the right, the write head edits the memory, writing information about the goals down. Ultimately, the DNC is commanded to satisfy goal \Q", which it does subsequently by using the read heads to inspect the locations containing goal Q. The constraints constituting goal \Q" are shown below, and the final board position is correct. (MOV 5292 kb)

PowerPoint slides

Rights and permissions

Reprints and permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Graves, A., Wayne, G., Reynolds, M. et al. Hybrid computing using a neural network with dynamic external memory. Nature 538, 471–476 (2016). https://doi.org/10.1038/nature20101

Download citation

  • Received:

  • Accepted:

  • Published:

  • Issue Date:

  • DOI: https://doi.org/10.1038/nature20101

This article is cited by

Comments

By submitting a comment you agree to abide by our Terms and Community Guidelines. If you find something abusive or that does not comply with our terms or guidelines please flag it as inappropriate.

Search

Quick links

Nature Briefing

Sign up for the Nature Briefing newsletter — what matters in science, free to your inbox daily.

Get the most important science stories of the day, free in your inbox. Sign up for Nature Briefing