Inferring computational dynamics from neural measurements using deep recurrent neural networks
Final Report Abstract
The present, mainly methodological project developed a novel class of deep learning algorithms for learning generative surrogate models from time series data, neurophysiological recordings in particular, an area called dynamical systems reconstruction. Such models are trained on data to mimic the dynamical properties of a system that has been empirically observed. After successful training, these models can then be used to generate predictions, or to analyse certain formal or computational properties of the system they have been trained on. At the time this research was originally conceived, this was still a nascent field, in which there was no established or well working approach. This project, in our minds, contributed to hugely advancing the field by identifying crucial challenges in training recurrent neural networks (RNNs), a class of deep learning models, on dynamical systems reconstruction problems. By identifying crucial mathematical issues in the training process, training algorithms and model architectures could be strongly improved. The developed methods were thoroughly tested and benchmarked on a variety of simulated systems and real-world datasets, exemplifying their universal applicability in many areas of science beyond neuroscience, for instance also in medical domains or in climate science. In neuroscience in particular, they can serve to dissect the computational processes by which the brain solves behavioural problems. While we did not progress as far as we originally intended in using these new AI tools for analysing neuronal recordings, we performed a number of other analyses that gave insight into the role of different brain regions in the dynamical mechanisms supporting working memory and decision making. Overall, this project established a new toolbox for the dynamical and computational analyses of neuronal systems.
Publications
-
Identifying nonlinear dynamical systems via generative recurrent neural networks with applications to fMRI. PLOS Computational Biology, 15(8), e1007263.
Koppe, Georgia; Toutounji, Hazem; Kirsch, Peter; Lis, Stefanie & Durstewitz, Daniel
-
Deep learning for small and big data in psychiatry. Neuropsychopharmacology, 46(1), 176-190.
Koppe, Georgia; Meyer-Lindenberg, Andreas & Durstewitz, Daniel
-
Existence of n-cycles and border-collision bifurcations in piecewise-linear continuous maps with applications to recurrent neural networks. Nonlinear Dynamics, 101(2), 1037-1052.
Monfared, Z. & Durstewitz, D.
-
Transformation of ReLU-based recurrent neural networks from discrete-time to continuous-time. in Proceedings of the 37th International Conference on Machine Learning (PMLR, 2020b). 119: p. 6999-7009
Monfared, Z. & D. Durstewitz
-
Brain network dynamics during working memory are modulated by dopamine and diminished in schizophrenia. Nature Communications, 12(1).
Braun, Urs; Harneit, Anais; Pergola, Giulio; Menara, Tommaso; Schäfer, Axel; Betzel, Richard F.; Zang, Zhenxiang; Schweiger, Janina I.; Zhang, Xiaolong; Schwarz, Kristina; Chen, Junfang; Blasi, Giuseppe; Bertolino, Alessandro; Durstewitz, Daniel; Pasqualetti, Fabio; Schwarz, Emanuel; Meyer-Lindenberg, Andreas; Bassett, Danielle S. & Tost, Heike
-
Classification of amyotrophic lateral sclerosis by brain volume, connectivity, and network dynamics. Human Brain Mapping, 43(2), 681-699.
Thome, Janine; Steinbach, Robert; Grosskreutz, Julian; Durstewitz, Daniel & Koppe, Georgia
-
Identifying nonlinear dynamical systems with multiple time scales and long-range dependencies. in Proceedings of the 9th International Conference on Learning Representations (ICLR, 2021)
Schmidt, D., G. Koppe, Z. Monfared, M. Beutelspacher & D. Durstewitz
-
Psychiatric Illnesses as Disorders of Network Dynamics. Biological Psychiatry: Cognitive Neuroscience and Neuroimaging, 6(9), 865-876.
Durstewitz, Daniel; Huys, Quentin J.M. & Koppe, Georgia
-
Conserved structures of neural activity in sensorimotor cortex of freely moving rats allow cross-subject decoding. Nature Communications, 13(1).
Melbaum, Svenja; Russo, Eleonora; Eriksson, David; Schneider, Artur; Durstewitz, Daniel; Brox, Thomas & Diester, Ilka
-
On the difficulty of learning chaotic dynamics with RNNs. in Proceedings of the 36th Annual Conference on Neural Information Processing Systems (NeurIPS, 2022). 35: p. 11297-11312
Mikhaeil, J., Z. Monfared & D. Durstewitz
-
Reconstructing Nonlinear Dynamical Systems from Multi-Modal Time Series. in Proceedings of the 39th International Conference on Machine Learning (PMLR, 2022). 162: p. 11613- 11633
Kramer, D., P.L. Bommer, C. Tombolini, G. Koppe & D. Durstewitz
-
Species-conserved mechanisms of abstract rule learning promote cognitive flexibility in complex environments. openRxiv.
Bähner, Florian; Popov, Tzvetan; Boehme, Nico; Hermann, Selina; Merten, Tom; Zingone, Hélène; Koppe, Georgia; Meyer-Lindenberg, Andreas; Toutounji, Hazem & Durstewitz, Daniel
-
Tractable Dendritic RNNs for Reconstructing Nonlinear Dynamical Systems. in Proceedings of the 39th International Conference on Machine Learning (PMLR, 2022). 162: p. 2292-2320
Brenner, M., F. Hess, J.M. Mikhaeil, L.F. Bereska, Z. Monfared, P.-C. Kuo & D. Durstewitz
-
Bifurcations and loss jumps in RNN training. in Proceedings of the 37th Annual Conference on Neural Information Processing Systems (NeurIPS, 2023). 36: p. 70511-70547
Eisenmann, L., Z. Monfared, N. Göring & D. Durstewitz
-
Distinct hippocampal-prefrontal neural assemblies coordinate memory encoding, maintenance, and recall. Current Biology, 33(7), 1220-1236.e4.
Domanski, Aleksander P.F.; Kucewicz, Michal T.; Russo, Eleonora; Tricklebank, Mark D.; Robinson, Emma S.J.; Durstewitz, Daniel & Jones, Matt W.
-
Generalized Teacher Forcing for Learning Chaotic Dynamics. in Proceedings of the 40th International Conference on Machine Learning (PMLR, 2023). 202: p. 13017-13049
Hess, F., Z. Monfared, M. Brenner & D. Durstewitz
-
Reconstructing computational system dynamics from neural data with recurrent neural networks. Nature Reviews Neuroscience, 24(11), 693-710.
Durstewitz, Daniel; Koppe, Georgia & Thurm, Max Ingo
-
Integration of rate and phase codes by hippocampal cell-assemblies supports flexible encoding of spatiotemporal context. Nature Communications, 15(1).
Russo, Eleonora; Becker, Nadine; Domanski, Aleks P. F.; Howe, Timothy; Freud, Kipp; Durstewitz, Daniel & Jones, Matthew W.
