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:

Learning efficient backprojections across cortical hierarchies in real time

A preprint version of the article is available at arXiv.

Abstract

Models of sensory processing and learning in the cortex need to efficiently assign credit to synapses in all areas. In deep learning, a known solution is error backpropagation, which requires biologically implausible weight transport from feed-forwards to feedback paths. We introduce phaseless alignment learning, a bio-plausible method to learn efficient feedback weights in layered cortical hierarchies. This is achieved by exploiting the noise naturally found in biophysical systems as an additional carrier of information. In our dynamical system, all weights are learned simultaneously with always-on plasticity and using only information locally available to the synapses. Our method is completely phase-free (no forwards and backwards passes or phased learning) and allows for efficient error propagation across multi-layer cortical hierarchies, while maintaining biologically plausible signal transport and learning. Our method is applicable to a wide class of models and improves on previously known biologically plausible ways of credit assignment: compared to random synaptic feedback, it can solve complex tasks with fewer neurons and learn more useful latent representations. We demonstrate this on various classification tasks using a cortical microcircuit model with prospective coding.

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

Fig. 1: Sensory processing over cortical hierarchies.
Fig. 2: Cortical microcircuit setup with one hidden layer.
Fig. 3: PAL aligns weight updates with BP in deep networks.
Fig. 4: PAL improves learning on teacher–student and classification tasks compared to fixed random synaptic feedback.
Fig. 5: PAL learns useful latent representations, where FA fails to do so.
Fig. 6: PAL outperforms (D)FA on CIFAR-10 using a continuous-time convolutional neural network.

Similar content being viewed by others

Data availability

The datasets analysed during the current study are publicly available in the following repositories: https://github.com/lkriener/yin_yang_data_set (Yin-Yang dataset40), http://yann.lecun.com/exdb/mnist/ (MNIST dataset71) and https://www.cs.toronto.edu/~kriz/cifar.html (CIFAR-10, ref. 72).

Code availability

The simulations used custom code written in Python using numpy (v.1.26.2) and PyTorch (v.2.0.1+cu117). Some simulations were made using a custom module for the GeNN simulation suite. All code is made available under https://zenodo.org/records/10405083 (ref. 73).

References

  1. Yamins, D. L. K. & DiCarlo, J. J. Using goal-driven deep learning models to understand sensory cortex. Nat. Neurosci. 19, 356–365 (2016).

  2. Richards, B. A. et al. A deep learning framework for neuroscience. Nat. Neurosci. 22, 1761–1770 (2019).

    Article  Google Scholar 

  3. Lillicrap, T. P., Santoro, A., Marris, L., Akerman, C. J. & Hinton, G. Backpropagation and the brain. Nat. Rev. Neurosci. 21, 335–346 (2020).

    Article  Google Scholar 

  4. Roelfsema, P. & Ooyen, A. Attention-gated reinforcement learning of internal representations for classification. Neural Comput. 17, 2176–2214 (2005).

    Article  Google Scholar 

  5. Costa, R. P., Assael, Y. M., Shillingford, B., de Freitas, N. & Vogels, T. P. Cortical microcircuits as gated-recurrent neural networks. In Proc. 31st Annual Conference on Neural Information Processing Systems (NIPS) (eds von Luxburg, U. et al.) 272–283 (ACM, 2017).

  6. Scellier, B. & Bengio, Y. Equilibrium propagation: bridging the gap between energy-based models and backpropagation. Front. Comput. Neurosci. 11, 24 (2017).

    Article  Google Scholar 

  7. Whittington, J. C. R. & Bogacz, R. An approximation of the error backpropagation algorithm in a predictive coding network with local Hebbian synaptic plasticity. Neural Comput. 29, 1229–1262 (2017).

    Article  MathSciNet  Google Scholar 

  8. Sacramento, J., Ponte Costa, R., Bengio, Y. & Senn, W. Dendritic cortical microcircuits approximate the backpropagation algorithm. In Proc. 31st Annual Conference on Neural Information Processing Systems (NIPS) (eds Bengio, S. et al.) 8721–8732 (ACM, 2018).

  9. Haider, P. et al. Latent equilibrium: a unified learning theory for arbitrarily fast computation with arbitrarily slow neurons. Adv. Neural Inf. Process. Syst. 34, 17839–17851 (2021).

    Google Scholar 

  10. Lillicrap, T. P., Cownden, D., Tweed, D. B. & Akerman, C. J. Random synaptic feedback weights support error backpropagation for deep learning. Nat. Commun. 7, 13276 (2016).

    Article  Google Scholar 

  11. Payeur, A., Guerguiev, J., Zenke, F., Richards, B. A. & Naud, R. Burst-dependent synaptic plasticity can coordinate learning in hierarchical circuits. Nat. Neurosci. 24, 1010–1019 (2021).

    Article  Google Scholar 

  12. Marblestone, A. H., Wayne, G. & Kording, K. P. Toward an integration of deep learning and neuroscience. Front. Comput. Neurosci. 10, 94 (2016).

    Article  Google Scholar 

  13. Haak, K. V. & Beckmann, C. F. Objective analysis of the topological organization of the human cortical visual connectome suggests three visual pathways. Cortex 98, 73–83 (2018).

    Article  Google Scholar 

  14. Friedrich, J., Urbanczik, R. & Senn, W. Spatio-temporal credit assignment in neuronal population learning. PLoS Comput. Biol. 7, e1002092 (2011).

    Article  MathSciNet  Google Scholar 

  15. Rumelhart, D. E., Hinton, G. E. & Williams, R. J. Learning representations by back-propagating errors. Nature 323, 533–536 (1986).

    Article  Google Scholar 

  16. LeCun, Y. A theoretical framework for back-propagation. In Proc. 1988 Connectionist Models Summer School (eds Touretzky, D. et al.) 21–28 (Morgan Kaufmann, 1988).

  17. Nøkland, A. Direct feedback alignment provides learning in deep neural networks. Adv. Neural Information Proc. Syst. 29, 1037–1045 (2016).

  18. Kolen, J. F. & Pollack, J. B. Backpropagation without weight transport. In Proc. 1994 IEEE International Conference on Neural Networks (ICNN’94) 1375–1380 (IEEE, 1994).

  19. Akrout, M., Wilson, C., Humphreys, P. C., Lillicrap, T. & Tweed, D. Deep learning without weight transport. Preprint at https://arxiv.org/abs/1904.05391 (2019).

  20. Lansdell, B. J., Prakash, P. R. & Kording, K. P. Learning to solve the credit assignment problem. In Proc. International Conference on Learning Representations (ICLR, 2020).

  21. Ernoult, M. M. et al. Towards scaling difference target propagation by learning backprop targets. In Proc. 39th International Conference on Machine Learning (eds Chaudhuri, K. et al.) 5968–5987 (ML Research Press, 2022).

  22. Bengio, Y. How auto-encoders could provide credit assignment in deep networks via target propagation. Preprint at https://arxiv.org/abs/1407.7906 (2014).

  23. Lee, D.-H., Zhang, S., Fischer, A. & Bengio, Y. Difference target propagation. In Proc. Joint European Conference on Machine Learning and Knowledge Discovery in Databases (eds Appice, A. et al.) 498–515 (Springer, 2015).

  24. Meulemans, A., Carzaniga, F., Suykens, J., Sacramento, J. & Grewe, B. F. A theoretical framework for target propagation. In Proc. Advances in Neural Information Processing Systems (eds Larochelle, H. et al.) 20024–20036 (Curran Associates, Inc., 2020).

  25. Meulemans, A. et al. Credit assignment in neural networks through deep feedback control. In Proc. 35th Conference on Advances in Neural Information Processing Systems (eds Ranzato, M. et al.) Vol. 34 (Curran Associates, Inc., 2021).

  26. O’Reilly, R. C. Biologically plausible error-driven learning using local activation differences: the generalized recirculation algorithm. Neural Comput. 8, 895–938 (1996).

    Article  Google Scholar 

  27. Ackley, D. H., Hinton, G. E. & Sejnowski, T. J. in Readings in Computer Vision: Issues, Problems, Principles, and Paradigms (eds Fischler, M. A. & Firschein, O.) 522–533 (Morgan Kaufmann, 1987).

  28. Bengio, Y. & Fischer, A. Early inference in energy-based models approximates back-propagation. Preprint at https://arxiv.org/abs/1510.02777 (2015).

  29. Guerguiev, J., Lillicrap, T. P. & Richards, B. A. Towards deep learning with segregated dendrites. eLife 6, e22901 (2017).

    Article  Google Scholar 

  30. Mesnard, T., Vignoud, G., Sacramento, J., Senn, W. & Bengio, Y. Ghost units yield biologically plausible backprop in deep neural networks. Preprint at https://arxiv.org/abs/1911.08585 (2019).

  31. Xie, X. & Seung, H. S. Equivalence of backpropagation and contrastive Hebbian learning in a layered network. Neural Comput. 15, 441–454 (2003).

    Article  Google Scholar 

  32. Song, Y. et al. Inferring neural activity before plasticity: a foundation for learning beyond backpropagation. Nat. Neurosci. 27, 348–358 (2022).

  33. Pozzi, I., Bohte, S. & Roelfsema, P. Attention-gated brain propagation: how the brain can implement reward-based error backpropagation. In Proc. Advances in Neural Information Processing Systems (eds Larochelle, H. et al.) 2516–2526 (Curran Associates, Inc., 2020).

  34. Pozzi, I., Bohté, S. & Roelfsema, P. A biologically plausible learning rule for deep learning in the brain. Preprint at https://arxiv.org/abs/1811.01768 (2018).

  35. Moskovitz, T. H., Litwin-Kumar, A. & Abbott, L. F. Feedback alignment in deep convolutional networks. Preprint at https://arxiv.org/abs/1812.06488 (2018).

  36. Bartunov, S. et al. Assessing the scalability of biologically-motivated deep learning algorithms and architectures. Preprint at https://arxiv.org/abs/1807.04587 (2018).

  37. Bidoret, C., Ayon, A., Barbour, B. & Casado, M. Presynaptic nr2a-containing NMDA receptors implement a high-pass filter synaptic plasticity rule. Proc. Natl Acad. Sci. USA 106, 14126–14131 (2009).

    Article  Google Scholar 

  38. Clopath, C., Büsing, L., Vasilaki, E. & Gerstner, W. Connectivity reflects coding: a model of voltage-based STDP with homeostasis. Nat. Neurosci. 13, 344–352 (2010).

  39. Bono, J. & Clopath, C. Modeling somatic and dendritic spike mediated plasticity at the single neuron and network level. Nat. Commun. 8, 706 (2017).

    Article  Google Scholar 

  40. Kriener, L., Göltz, J. & Petrovici, M. A. The Yin-Yang dataset. Preprint at https://arxiv.org/abs/2102.08211 (2021).

  41. Greedy, W., Zhu, H. W., Pemberton, J., Mellor, J. & Ponte Costa, R. Single-phase deep learning in cortico-cortical networks. In Proc. Advances in Neural Information Processing Systems (eds Koyejo, S. et al.) 24213–24225 (NeurIPS, 2022).

  42. Crafton, B., Parihar, A., Gebhardt, E. & Raychowdhury, A. Direct feedback alignment with sparse connections for local learning. Front. Neurosci. 13, 525 (2019).

    Article  Google Scholar 

  43. Sato, H., Hata, Y., Masui, H. & Tsumoto, T. A functional role of cholinergic innervation to neurons in the cat visual cortex. J. Neurophysiol. 58, 765–780 (1987).

    Article  Google Scholar 

  44. Soma, S., Shimegi, S., Osaki, H. & Sato, H. Cholinergic modulation of response gain in the primary visual cortex of the macaque. J. Neurophysiol. 107, 283–291 (2012).

    Article  Google Scholar 

  45. Kang, J. I., Huppé-Gourgues, F. & Vaucher, E. Boosting visual cortex function and plasticity with acetylcholine to enhance visual perception. Front. Syst. Neurosci. 8, 172 (2014).

  46. Cornford, J. et al. Learning to live with Dale’s principle: ANNs with separate excitatory and inhibitory units. Preprint at bioRxiv https://doi.org/10.1101/2020.11.02.364968 (2021).

  47. Burbank, K. S. Mirrored STDP implements autoencoder learning in a network of spiking neurons. PLoS Comput. Biol. 11, e1004566 (2015).

    Article  Google Scholar 

  48. Maass, W. Noise as a resource for computation and learning in networks of spiking neurons. Proc. IEEE 102, 860–880 (2014).

    Article  Google Scholar 

  49. Rusakov, D. A., Savtchenko, L. P. & Latham, P. E. Noisy synaptic conductance: bug or a feature? Trends Neurosci. 43, 363–372 (2020).

    Article  Google Scholar 

  50. McDonnell, M. D. & Ward, L. M. The benefits of noise in neural systems: bridging theory and experiment. Nat. Rev. Neurosci. 12, 415–425 (2011).

    Article  Google Scholar 

  51. Faisal, A. A., Selen, L. P. J. & Wolpert, D. M. Noise in the nervous system. Nat. Rev. Neurosci. 9, 292–303 (2008).

    Article  Google Scholar 

  52. Xie, X. & Seung, H. S. Learning in neural networks by reinforcement of irregular spiking. Phys. Rev. E 69, 041909 (2004).

  53. Fiete, I. R., Fee, M. S. & Seung, H. S. Model of birdsong learning based on gradient estimation by dynamic perturbation of neural conductances. J. Neurophysiol. 98, 2038–2057 (2007).

    Article  Google Scholar 

  54. Plesser, H. E. & Gerstner, W. Escape rate models for noisy integrate-and-free neurons. Neurocomputing 32, 219–224 (2000).

    Article  Google Scholar 

  55. Köndgen, H. et al. The dynamical response properties of neocortical neurons to temporally modulated noisy inputs in vitro. Cereb. Cortex 18, 2086–2097 (2008).

    Article  Google Scholar 

  56. Petrovici, M. A., Bill, J., Bytschok, I., Schemmel, J. & Meier, K. Stochastic inference with spiking neurons in the high-conductance state. Phys. Rev. E 94, 042312 (2016).

    Article  MathSciNet  Google Scholar 

  57. Ricciardi, L. M. & Sacerdote, L. The Ornstein–Uhlenbeck process as a model for neuronal activity. I. Mean and variance of the firing time. Biol. Cybern. 35, 1–9 (1979).

    Article  Google Scholar 

  58. Gerstner, W., Kistler, W. M., Naud, R. & Paninski, L. Neuronal Dynamics: From Single Neurons to Networks and Models of Cognition (Cambridge Univ. Press, 2014).

  59. Petrovici, M. A. Form Versus Function: Theory and Models for Neuronal Substrates Vol. 1 (Springer, 2016).

  60. Jordan, J., Sacramento, J., Wybo, W. A. M., Petrovici, M. A. & Senn, W. Learning Bayes-optimal dendritic opinion pooling. Preprint at https://arxiv.org/abs/2104.13238 (2021).

  61. Crochet, S., Poulet, J. F. A., Kremer, Y. & Petersen, C. C. H. Synaptic mechanisms underlying sparse coding of active touch. Neuron 69, 1160–1175 (2011).

    Article  Google Scholar 

  62. Szendro, P., Vincze, G. & Szasz, A. Bio-response to white noise excitation. Electro. Magnetobiol. 20, 215–229 (2001).

    Article  Google Scholar 

  63. Urbanczik, R. & Senn, W. Learning by the dendritic prediction of somatic spiking. Neuron 81, 521–528 (2014).

    Article  Google Scholar 

  64. Gerstner, W., Lehmann, M., Liakoni, V., Corneil, D. & Brea, J. Eligibility traces and plasticity on behavioral time scales: experimental support of neoHebbian three-factor learning rules. Front. Neural Circuits 12, 53 (2018).

  65. Jordan, R. & Keller, G. B. Opposing influence of top-down and bottom-up input on excitatory layer 2/3 neurons in mouse primary visual cortex. Neuron 108, 1194–1206 (2020).

    Article  Google Scholar 

  66. Körding, K. P. & König, P. Supervised and unsupervised learning with two sites of synaptic integration. J. Comput. Neurosci. 11, 207–215 (2001).

    Article  Google Scholar 

  67. Spruston, N. Pyramidal neurons: dendritic structure and synaptic integration. Nat. Rev. Neurosci. 9, 206–221 (2008).

  68. Särkkä, S. & Solin, A. Applied Stochastic Differential Equations (Cambridge Univ. Press, 2019).

  69. Yavuz, E., Turner, J. & Nowotny, T. GeNN: a code generation framework for accelerated brain simulations. Sci. Rep. 6, 18854 (2016).

    Article  Google Scholar 

  70. Knight, J. C., Komissarov, A. & Nowotny, T. PyGeNN: a python library for GPU-enhanced neural networks. Front. Neuroinform. 15, 659005 (2021).

    Article  Google Scholar 

  71. LeCun, Y., Bottou, L., Bengio, Y. & Haffner, P. Gradient-based learning applied to document recognition. Proc. IEEE 86, 2278–2324 (1998).

    Article  Google Scholar 

  72. Krizhevsky, A. et al. Learning Multiple Layers of Features from Tiny Images (Univ. Toronto, 2009).

  73. Max, K., Kriener, L. & Jaras, I. Code repository for phaseless alignment learning. Zenodo https://doi.org/10.5281/zenodo.10405083 (2024).

  74. Archer, K., Pammer, K. & Vidyasagar, T. R. A temporal sampling basis for visual processing in developmental dyslexia. Front. Hum. Neurosci. 14, 213 (2020).

    Article  Google Scholar 

  75. Gray, H. Anatomy of the Human Body (Lea & Febiger, 1918).

Download references

Acknowledgements

We thank J. Jordan, A. Meulemans and J. Sacramento for valuable discussions. We gratefully acknowledge funding from the European Union under grant agreement nos. 604102, 720270, 785907 and 945539 (Human Brain Project) and the Manfred Stärk Foundation. Additionally, our work has greatly benefited from access to the Fenix Infrastructure resources, which are partially funded from the European Union’s Horizon 2020 Research and Innovation programme through the ICEI project under the grant agreement no. 800858. This includes access to Piz Daint at the Swiss National Supercomputing Centre, Switzerland. Further calculations were performed on UBELIX, the High Performance Computing cluster at the University of Bern.

Author information

Authors and Affiliations

Authors

Contributions

K.M. derived, with contributions by L.K. and M.A.P., the PAL algorithm. K.M. and L.K. adapted the dendritic microcircuit model to include PAL for learning the feedback weights. G.P.G. and T.N. developed a dendritic microcircuit module for the GeNN simulator. L.K. added the latent equilibrium and PAL mechanisms to the module. K.M. and L.K. performed the simulation experiments. I.J. and K.M. worked on scaling the algorithm to a larger benchmark during the revision process. The paper was mainly written by K.M., aided by L.K. and M.A.P. M.A.P. and W.S. provided supervision and funding to this project.

Corresponding author

Correspondence to Kevin Max.

Ethics declarations

Competing interests

The authors declare no competing interests.

Peer review

Peer review information

Nature Machine Intelligence thanks Ben Lansdell and the other, anonymous, reviewer(s) for their contribution to the peer review of this work.

Additional information

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Supplementary information

Supplementary Information

Derivations, Extensions of theory, Supplementary Figs. 1–3, Algorithm 1 and Tables 1 and 2.

Reporting Summary

Rights and permissions

Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law.

Reprints and permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Max, K., Kriener, L., Pineda García, G. et al. Learning efficient backprojections across cortical hierarchies in real time. Nat Mach Intell 6, 619–630 (2024). https://doi.org/10.1038/s42256-024-00845-3

Download citation

  • Received:

  • Accepted:

  • Published:

  • Issue Date:

  • DOI: https://doi.org/10.1038/s42256-024-00845-3

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