Abstract representations emerge naturally in neural networks trained to perform multiple tasks
- PMID: 36823136
- PMCID: PMC9950464
- DOI: 10.1038/s41467-023-36583-0
Abstract representations emerge naturally in neural networks trained to perform multiple tasks
Abstract
Humans and other animals demonstrate a remarkable ability to generalize knowledge across distinct contexts and objects during natural behavior. We posit that this ability to generalize arises from a specific representational geometry, that we call abstract and that is referred to as disentangled in machine learning. These abstract representations have been observed in recent neurophysiological studies. However, it is unknown how they emerge. Here, using feedforward neural networks, we demonstrate that the learning of multiple tasks causes abstract representations to emerge, using both supervised and reinforcement learning. We show that these abstract representations enable few-sample learning and reliable generalization on novel tasks. We conclude that abstract representations of sensory and cognitive variables may emerge from the multiple behaviors that animals exhibit in the natural world, and, as a consequence, could be pervasive in high-level brain regions. We also make several specific predictions about which variables will be represented abstractly.
© 2023. The Author(s).
Conflict of interest statement
The authors declare no competing interests.
Figures
![Fig. 1](https://cdn.statically.io/img/www.ncbi.nlm.nih.gov/pmc/articles/instance/9950464/bin/41467_2023_36583_Fig1_HTML.gif)
![Fig. 2](https://cdn.statically.io/img/www.ncbi.nlm.nih.gov/pmc/articles/instance/9950464/bin/41467_2023_36583_Fig2_HTML.gif)
![Fig. 3](https://cdn.statically.io/img/www.ncbi.nlm.nih.gov/pmc/articles/instance/9950464/bin/41467_2023_36583_Fig3_HTML.gif)
![Fig. 4](https://cdn.statically.io/img/www.ncbi.nlm.nih.gov/pmc/articles/instance/9950464/bin/41467_2023_36583_Fig4_HTML.gif)
![Fig. 5](https://cdn.statically.io/img/www.ncbi.nlm.nih.gov/pmc/articles/instance/9950464/bin/41467_2023_36583_Fig5_HTML.gif)
![Fig. 6](https://cdn.statically.io/img/www.ncbi.nlm.nih.gov/pmc/articles/instance/9950464/bin/41467_2023_36583_Fig6_HTML.gif)
![Fig. 7](https://cdn.statically.io/img/www.ncbi.nlm.nih.gov/pmc/articles/instance/9950464/bin/41467_2023_36583_Fig7_HTML.gif)
Comment in
-
Unveiling the benefits of multitasking in disentangled representation formation.Trends Cogn Sci. 2023 Aug;27(8):699-701. doi: 10.1016/j.tics.2023.05.010. Epub 2023 Jun 23. Trends Cogn Sci. 2023. PMID: 37357063
Similar articles
-
Abstract representations emerge in human hippocampal neurons during inference behavior.bioRxiv [Preprint]. 2023 Nov 30:2023.11.10.566490. doi: 10.1101/2023.11.10.566490. bioRxiv. 2023. PMID: 37986878 Free PMC article. Preprint.
-
Reward-predictive representations generalize across tasks in reinforcement learning.PLoS Comput Biol. 2020 Oct 15;16(10):e1008317. doi: 10.1371/journal.pcbi.1008317. eCollection 2020 Oct. PLoS Comput Biol. 2020. PMID: 33057329 Free PMC article.
-
Neural representational geometry underlies few-shot concept learning.Proc Natl Acad Sci U S A. 2022 Oct 25;119(43):e2200800119. doi: 10.1073/pnas.2200800119. Epub 2022 Oct 17. Proc Natl Acad Sci U S A. 2022. PMID: 36251997 Free PMC article.
-
Representations and generalization in artificial and brain neural networks.Proc Natl Acad Sci U S A. 2024 Jul 2;121(27):e2311805121. doi: 10.1073/pnas.2311805121. Epub 2024 Jun 24. Proc Natl Acad Sci U S A. 2024. PMID: 38913896 Free PMC article. Review.
-
Continual task learning in natural and artificial agents.Trends Neurosci. 2023 Mar;46(3):199-210. doi: 10.1016/j.tins.2022.12.006. Epub 2023 Jan 20. Trends Neurosci. 2023. PMID: 36682991 Free PMC article. Review.
Cited by
-
Sequential neuronal processing of number values, abstract decision, and action in the primate prefrontal cortex.PLoS Biol. 2024 Feb 16;22(2):e3002520. doi: 10.1371/journal.pbio.3002520. eCollection 2024 Feb. PLoS Biol. 2024. PMID: 38364194 Free PMC article.
-
Improving reduced-order models through nonlinear decoding of projection-dependent outputs.Patterns (N Y). 2023 Oct 10;4(11):100859. doi: 10.1016/j.patter.2023.100859. eCollection 2023 Nov 10. Patterns (N Y). 2023. PMID: 38035196 Free PMC article.
-
Hierarchical VAEs provide a normative account of motion processing in the primate brain.bioRxiv [Preprint]. 2023 Nov 5:2023.09.27.559646. doi: 10.1101/2023.09.27.559646. bioRxiv. 2023. PMID: 37808629 Free PMC article. Preprint.
-
Reconstructing computational system dynamics from neural data with recurrent neural networks.Nat Rev Neurosci. 2023 Nov;24(11):693-710. doi: 10.1038/s41583-023-00740-7. Epub 2023 Oct 4. Nat Rev Neurosci. 2023. PMID: 37794121 Review.
-
A view-based decision mechanism for rewards in the primate amygdala.Neuron. 2023 Dec 6;111(23):3871-3884.e14. doi: 10.1016/j.neuron.2023.08.024. Epub 2023 Sep 18. Neuron. 2023. PMID: 37725980 Free PMC article.
References
Publication types
MeSH terms
Associated data
LinkOut - more resources
Full Text Sources
Miscellaneous