Reconstructing computational system dynamics from neural data with recurrent neural networks

The controllosphere: The neural origin of cognitive effort.

“…For example, the monkey brain–computer interface experiment conducted by Sadtler et al (2014; see “The Intrinsic Manifold”) could be replicated in humans using magnetoencephalography (e.g., Corsi et al, 2021; Stiso et al, 2020; Youssofzadeh et al, 2023). RNNs reproducing the intrinsic manifold could then either be constructed by hand (e.g., Wärnberg & Kumar, 2019) or reverse-engineered directly from the neural data (Durstewitz et al, 2023; see Box 3). The connectivity matrices of these networks could then be used to derive Gramians that formalize the shape of the intrinsic manifold.…”

“…However, this empirical approach requires that the experimenter set the initial state of the system and then apply impulsive inputs to it (Lall et al, 2002), which are difficult criteria to satisfy in (noninvasive) human experiments. A more promising approach is to train computational models directly on neural data in order to extract their underlying generative structure (Durstewitz, 2017; Durstewitz et al, 2023; Langdon & Engel, 2022; Pandarinath et al, 2018). For example, the intrinsic manifold can be produced by RNNs trained on fMRI data (Koppe et al, 2019) and/or task-related behavioral data (Z.…”

Trained recurrent neural networks develop phase-locked limit cycles in a working memory task

“…We used trained recurrent neural networks [20, 22, 24]. RNNs can be trained to reproduce both neural activity and behavior [17, 23, 31, 34, 36].…”

“…Specifically, we seek to characterize dynamical systems that could underlie working memory relying on phase coding with neural oscillations. We do so by assuming that cognitive functions can be described by a low-dimensional dynamical system, implemented through populations of neurons (computation through dynamics) [20][21][22][23][24][25][26], in line with empirical observations [27,28].…”

“…RNNs are universal dynamical systems, in the sense that they can approximate finite time trajectories of any dynamical system [29,30]. We are thus able to use these networks as hypothesis generators -first training them on a cognitive task, and then reverse engineering the resulting networks [17,20,24,26,[31][32][33][34][35][36][37][38].…”