Neural population dynamics of computing with synaptic modulations
Abstract
In addition to long-timescale rewiring, synapses in the brain are subject to significant modulation that occurs at faster timescales. These modulations vary widely in underlying biological mechanisms as well as the timescales over which they occur, yet they all endow the brain with additional means of processing information. Despite this, models of the brain like recurrent neural networks (RNNs) often have their weights frozen after training, relying on an internal state stored in neuronal activity to hold temporal information over task-relevant timescales. Although networks with dynamical synapses have been explored previously, often said modulations are added to networks that also have recurrent connections and thus the computational capabilities and dynamical behavior contributed by the synapses remain unclear. In this work, we study the computational potential and resulting dynamics of a network that relies solely on synapse dynamics to process temporal information, the multi-plasticity network (MPN). Unlike traditional RNNs, the weights in the MPN are modulated during inference. The generality of the MPN allows for our results to apply to synaptic modulation mechanisms ranging from short-term synaptic plasticity (STSP) to slower modulations such as spike-time dependent plasticity (STDP). We thoroughly examine the neural population dynamics of the MPN trained on integration-based tasks and compare it to known RNN dynamics, finding the two to have fundamentally different attractor structure. We find said differences in dynamics allow the MPN to outperform its RNN counterparts on several neuroscience-relevant tests. Training the MPN across a battery of neuroscience tasks, we find its computational capabilities in such settings is comparable to networks that compute with recurrent connections. Altogether, we believe this works demonstrates the computational possibilities of computing with synaptic modulations and highlights important motifs of these computations so that they can be identified in brain-like systems.
Data availability
The current manuscript is a computational study, so no data have been generated for this manuscript. Modeling code is available on GitHub at: https://github.com/kaitken17/mpn (the link provided in the Methods section of the paper).
Article and author information
Author details
Funding
National Institutes of Health (F1DA055669)
- Stefan Mihalas
National Institutes of Health (R01EB02981)
- Stefan Mihalas
The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.
Copyright
© 2023, Aitken & Mihalas
This article is distributed under the terms of the Creative Commons Attribution License permitting unrestricted use and redistribution provided that the original author and source are credited.
Metrics
-
- 1,850
- views
-
- 309
- downloads
-
- 6
- citations
Views, downloads and citations are aggregated across all versions of this paper published by eLife.
Download links
Downloads (link to download the article as PDF)
Open citations (links to open the citations from this article in various online reference manager services)
Cite this article (links to download the citations from this article in formats compatible with various reference manager tools)
Further reading
-
- Neuroscience
Non-linear summation of synaptic inputs to the dendrites of pyramidal neurons has been proposed to increase the computation capacity of neurons through coincidence detection, signal amplification, and additional logic operations such as XOR. Supralinear dendritic integration has been documented extensively in principal neurons, mediated by several voltage-dependent conductances. It has also been reported in parvalbumin-positive hippocampal basket cells, in dendrites innervated by feedback excitatory synapses. Whether other interneurons, which support feed-forward or feedback inhibition of principal neuron dendrites, also exhibit local non-linear integration of synaptic excitation is not known. Here, we use patch-clamp electrophysiology, and two-photon calcium imaging and glutamate uncaging, to show that supralinear dendritic integration of near-synchronous spatially clustered glutamate-receptor mediated depolarization occurs in NDNF-positive neurogliaform cells and oriens-lacunosum moleculare interneurons in the mouse hippocampus. Supralinear summation was detected via recordings of somatic depolarizations elicited by uncaging of glutamate on dendritic fragments, and, in neurogliaform cells, by concurrent imaging of dendritic calcium transients. Supralinearity was abolished by blocking NMDA receptors (NMDARs) but resisted blockade of voltage-gated sodium channels. Blocking L-type calcium channels abolished supralinear calcium signalling but only had a minor effect on voltage supralinearity. Dendritic boosting of spatially clustered synaptic signals argues for previously unappreciated computational complexity in dendrite-projecting inhibitory cells of the hippocampus.
-
- Neuroscience
Complex structural and functional changes occurring in typical and atypical development necessitate multidimensional approaches to better understand the risk of developing psychopathology. Here, we simultaneously examined structural and functional brain network patterns in relation to dimensions of psychopathology in the Adolescent Brain Cognitive Development (ABCD) dataset. Several components were identified, recapitulating the psychopathology hierarchy, with the general psychopathology (p) factor explaining most covariance with multimodal imaging features, while the internalizing, externalizing, and neurodevelopmental dimensions were each associated with distinct morphological and functional connectivity signatures. Connectivity signatures associated with the p factor and neurodevelopmental dimensions followed the sensory-to-transmodal axis of cortical organization, which is related to the emergence of complex cognition and risk for psychopathology. Results were consistent in two separate data subsamples and robust to variations in analytical parameters. Although model parameters yielded statistically significant brain–behavior associations in unseen data, generalizability of the model was rather limited for all three latent components (r change from within- to out-of-sample statistics: LC1within = 0.36, LC1out = 0.03; LC2within = 0.34, LC2out = 0.05; LC3within = 0.35, LC3out = 0.07). Our findings help in better understanding biological mechanisms underpinning dimensions of psychopathology, and could provide brain-based vulnerability markers.