Richard Kempter
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Richard Kempter has a total of 122 co-authors. Here are the top co-authors by number of shared publications:
- Paula Tuulia Kuokkanen
Institute for Theoretical Biology, Humboldt-Universität zu Berlin, 10115 Berlin, Germany
10 shared publications - Dietmar Schmitz
Neuroscience Research Center, Charite-Universitaetsmedizin Berlin, Germany
9 shared publications - Catherine Emily Carr
Department of Biology, University of Maryland College Park
8 shared publications - Nikolaus Maier
Charité-Universitätsmedizin Berlin, corporate member of Freie Universität Berlin and Humboldt-Universität zu Berlin, Neuroscience Research Center
7 shared publications - Gaspar Cano
Institute for Theoretical Biology, Department of Biology, Humboldt-Universität zu Berlin
5 shared publications - Naomi Auer
Affiliation Institute for Theoretical Biology, Department of Biology, Humboldt-Universität zu Berlin, Berlin, Germany
5 shared publications
Latest Publications
- Open Access
- 6 July 2026
- Published by Public Library of Science (PLoS) in PLOS Computational Biology
- Vol. 22 (7) , e1013235
Population sparseness determines strength of Hebbian plasticity for maximal memory lifetime in associative networks
The brain can efficiently learn and form memories based on limited exposure to stimuli, often even in single trials. Two key factors are believed to support this ability: large synaptic plasticity to strongly encode new memories; and sparse coding, leading to low overlap between memory representations and to small interference. Therefore, increased sparseness can also improve memory capacity. However, it is not well understood how the strength of plasticity of synapses affects capacity. Here, we analyze the combined impact of population sparseness and strength of plasticity on memory capacity. Specifically, we explore how the strength of plasticity that maximizes capacity depends on the sparseness of the neural code. To this end, we study a feedforward network with Hebbian and homeostatic plasticity and a two-state synapse model. The network learns to associate sparse binary input-output pattern pairs. The strength of plasticity is modeled as the probability of synaptic changes. Our results are based on both network simulations and an analytical theory, predicting the expected memory capacity in dependence on strength of plasticity and population sparseness. For both perfect and noisy input patterns, we find that the optimal strength of plasticity increases with increasing pattern sparseness and that this effect is more pronounced for input than for output sparseness. Interestingly, the optimal strength of plasticity remains the same across different network sizes if the number of active units in an input pattern is constant. While the memory capacity obtained at the optimal strength of plasticity increases monotonically with output sparseness, its dependence on input sparseness is non-monotonic. Overall, we provide the first detailed investigation of the interactions between population sparseness, strength of plasticity, and memory capacity. Our findings suggest that differences in sparseness between brain regions may underlie observed differences in how strongly these regions adapt and how quickly they learn.
- 6 July 2026
- Published by Public Library of Science (PLoS) in PLOS Computational Biology
- Vol. 22 (7) , e1013235
Open AccessPopulation sparseness determines strength of Hebbian plasticity for maximal memory lifetime in associative networks
The brain can efficiently learn and form memories based on limited exposure to stimuli, often even in single trials. Two key factors are believed to support this ability: large synaptic plasticity to strongly encode new memories; and sparse coding, leading to low overlap between memory representations and to small interference. Therefore, increased sparseness can also improve memory capacity. However, it is not well understood how the strength of plasticity of synapses affects capacity. Here, we analyze the combined impact of population sparseness and strength of plasticity on memory capacity. Specifically, we explore how the strength of plasticity that maximizes capacity depends on the sparseness of the neural code. To this end, we study a feedforward network with Hebbian and homeostatic plasticity and a two-state synapse model. The network learns to associate sparse binary input-output pattern pairs. The strength of plasticity is modeled as the probability of synaptic changes. Our results are based on both network simulations and an analytical theory, predicting the expected memory capacity in dependence on strength of plasticity and population sparseness. For both perfect and noisy input patterns, we find that the optimal strength of plasticity increases with increasing pattern sparseness and that this effect is more pronounced for input than for output sparseness. Interestingly, the optimal strength of plasticity remains the same across different network sizes if the number of active units in an input pattern is constant. While the memory capacity obtained at the optimal strength of plasticity increases monotonically with output sparseness, its dependence on input sparseness is non-monotonic. Overall, we provide the first detailed investigation of the interactions between population sparseness, strength of plasticity, and memory capacity. Our findings suggest that differences in sparseness between brain regions may underlie observed differences in how strongly these regions adapt and how quickly they learn.
- Open Access
Elementary Dynamics of Neural Microcircuits
In neural circuits with multiple excitatory (E) and inhibitory (I) cell types, qualitatively new dynamics arise when one population is silenced by the others, an effect linear theories cannot capture. Using threshold-linear networks, the simplest nonlinear population model, we derive closed-form conditions for multistability, oscillations, and paradoxical responses across microcircuits. In E-I circuits, strengthening E → I rather than I → E connections controls runaway excitation without silencing E cells. In a canonical neocortical circuit, two distinct gamma-oscillation mechanisms reproduce optogenetic silencing experiments, reconciling conflicting models. In hippocampal E-I-I circuits, cross-inhibition determines whether transitions between gamma rhythms are gradual, abrupt, or non-oscillatory. Two E populations or assemblies can cooperate, compete, or suppress each other depending on cross-versus self-excitation. Balanced inhibition prevents competition between E-I assemblies, unless I → I connections are more specialized than I → E ones, whereas lateral inhibition yields winner-take-all dynamics even without assemblies. These testable conditions link connectivity to dynamics across brain areas.
Open AccessElementary Dynamics of Neural Microcircuits
In neural circuits with multiple excitatory (E) and inhibitory (I) cell types, qualitatively new dynamics arise when one population is silenced by the others, an effect linear theories cannot capture. Using threshold-linear networks, the simplest nonlinear population model, we derive closed-form conditions for multistability, oscillations, and paradoxical responses across microcircuits. In E-I circuits, strengthening E → I rather than I → E connections controls runaway excitation without silencing E cells. In a canonical neocortical circuit, two distinct gamma-oscillation mechanisms reproduce optogenetic silencing experiments, reconciling conflicting models. In hippocampal E-I-I circuits, cross-inhibition determines whether transitions between gamma rhythms are gradual, abrupt, or non-oscillatory. Two E populations or assemblies can cooperate, compete, or suppress each other depending on cross-versus self-excitation. Balanced inhibition prevents competition between E-I assemblies, unless I → I connections are more specialized than I → E ones, whereas lateral inhibition yields winner-take-all dynamics even without assemblies. These testable conditions link connectivity to dynamics across brain areas.
- Open Access
- 13 April 2026
- Published by American Physical Society (APS) in PRX Life
- Vol. 4 (2)
Stochastic Synaptic Dynamics Under Learning
Learning is based on synaptic plasticity, which affects and is driven by neural activity. Because pre- and postsynaptic spiking activity is shaped by randomness, the synaptic weights follow a stochastic process, requiring a probabilistic framework to capture the noisy synaptic dynamics. We consider a paradigmatic supervised learning example: a presynaptic neural population impinging in a sequence of episodes on a recurrent network of integrate-and-fire neurons through synapses undergoing spike-timing-dependent plasticity (STDP) with additive potentiation and multiplicative depression. We first analytically compute the drift and diffusion coefficients for a single synapse within a single episode (microscopic dynamics), mapping the true jump process to a Langevin and the associated Fokker-Planck equations. Leveraging recent analytical tools, we include spike-time-resolving cross-correlations between pre- and postsynaptic spikes, which corrects substantial deviations seen in standard theories purely based on firing rates. We then apply this microdynamical description to the network setup in which hetero-associations are trained over one-shot episodes into a feed-forward matrix of STDP synapses connecting to neurons of the recurrent network (macroscopic dynamics). By mapping statistically distinct synaptic populations to instances of the single-synapse process above, we self-consistently determine the joint neural and synaptic dynamics and, ultimately, the time course of memory degradation and the memory capacity. We demonstrate that specifically in the relevant case of sparse coding, our theory can quantitatively capture memory capacities which are strongly overestimated if spike-time–resolving cross-correlations are ignored. On the single-synapse level, our theory predicts a transition between acceleration and deceleration of the synaptic weight, and a possibly nonmonotonic dependence of the finite-time diffusion coefficient on the synaptic weight. On the memory level, we showcase two complementary roles of homeostatic plasticity: namely, it is both responsible for forgetting and a requirement for the ability to learn. Additionally, we show how strength of plasticity and duration of exposure to stimuli must be scaled to evade a tradeoff dilemma between memory capacity and early accuracy. We conclude with a discussion of the many directions in which our framework can be extended.
- 13 April 2026
- Published by American Physical Society (APS) in PRX Life
- Vol. 4 (2)
Open AccessStochastic Synaptic Dynamics Under Learning
Learning is based on synaptic plasticity, which affects and is driven by neural activity. Because pre- and postsynaptic spiking activity is shaped by randomness, the synaptic weights follow a stochastic process, requiring a probabilistic framework to capture the noisy synaptic dynamics. We consider a paradigmatic supervised learning example: a presynaptic neural population impinging in a sequence of episodes on a recurrent network of integrate-and-fire neurons through synapses undergoing spike-timing-dependent plasticity (STDP) with additive potentiation and multiplicative depression. We first analytically compute the drift and diffusion coefficients for a single synapse within a single episode (microscopic dynamics), mapping the true jump process to a Langevin and the associated Fokker-Planck equations. Leveraging recent analytical tools, we include spike-time-resolving cross-correlations between pre- and postsynaptic spikes, which corrects substantial deviations seen in standard theories purely based on firing rates. We then apply this microdynamical description to the network setup in which hetero-associations are trained over one-shot episodes into a feed-forward matrix of STDP synapses connecting to neurons of the recurrent network (macroscopic dynamics). By mapping statistically distinct synaptic populations to instances of the single-synapse process above, we self-consistently determine the joint neural and synaptic dynamics and, ultimately, the time course of memory degradation and the memory capacity. We demonstrate that specifically in the relevant case of sparse coding, our theory can quantitatively capture memory capacities which are strongly overestimated if spike-time–resolving cross-correlations are ignored. On the single-synapse level, our theory predicts a transition between acceleration and deceleration of the synaptic weight, and a possibly nonmonotonic dependence of the finite-time diffusion coefficient on the synaptic weight. On the memory level, we showcase two complementary roles of homeostatic plasticity: namely, it is both responsible for forgetting and a requirement for the ability to learn. Additionally, we show how strength of plasticity and duration of exposure to stimuli must be scaled to evade a tradeoff dilemma between memory capacity and early accuracy. We conclude with a discussion of the many directions in which our framework can be extended.
- Open Access
- 16 January 2026
- Published by Public Library of Science (PLoS) in PLOS Computational Biology
- Vol. 22 (1) , e1013844
Conditions for replay of neuronal assemblies
From cortical synfire chains to hippocampal replay, the idea that neural populations can be activated sequentially with precise spike timing is thought to be essential for several brain functions. It has been shown that neuronal sequences with weak feedforward connectivity can be replayed due to amplification via intra-assembly recurrent connections. However, the mechanisms behind this phenomenon are still unclear. Here, we simulate spiking networks with different excitatory and inhibitory connectivity and find that an exclusively excitatory network is sufficient for this amplification to occur. To explain the spiking network behavior, we introduce a population model of membrane-potential distributions, and we analytically describe how different connectivity structures determine replay speed, with weaker feedforward connectivity generating slower and wider pulses that can be sustained by recurrent connections. Pulse propagation is facilitated if the neuronal membrane time constant is large compared to the pulse width. Together, our simulations and analytical results predict the conditions for replay of neuronal assemblies.
- 16 January 2026
- Published by Public Library of Science (PLoS) in PLOS Computational Biology
- Vol. 22 (1) , e1013844
Open AccessConditions for replay of neuronal assemblies
From cortical synfire chains to hippocampal replay, the idea that neural populations can be activated sequentially with precise spike timing is thought to be essential for several brain functions. It has been shown that neuronal sequences with weak feedforward connectivity can be replayed due to amplification via intra-assembly recurrent connections. However, the mechanisms behind this phenomenon are still unclear. Here, we simulate spiking networks with different excitatory and inhibitory connectivity and find that an exclusively excitatory network is sufficient for this amplification to occur. To explain the spiking network behavior, we introduce a population model of membrane-potential distributions, and we analytically describe how different connectivity structures determine replay speed, with weaker feedforward connectivity generating slower and wider pulses that can be sustained by recurrent connections. Pulse propagation is facilitated if the neuronal membrane time constant is large compared to the pulse width. Together, our simulations and analytical results predict the conditions for replay of neuronal assemblies.
- Open Access
- 1 January 2026
- Published by SAGE Publications in Trends in Hearing
- Vol. 30
Inhibitory inputs to avian ITD circuits
In birds, detection of sound source azimuth begins in the nucleus laminaris (NL), which computes interaural time differences (ITDs). Inhibition has been proposed to protect NL neurons from losing ITD sensitivity with increasing sound level, although the nature of this inhibition is incompletely understood. Anatomical studies show GABAergic synapses throughout NL and first order nuclei, while in vitro studies in chicken NL showed that GABAergic inhibitory postsynaptic potentials (IPSPs) shorten membrane time constants, perhaps to allow the membrane potential to follow rapid synaptic currents accurately over a range of sound levels. Given the importance of inhibition in regulating auditory brainstem activity, we examined the nature of inhibitory input to NL in vivo. The Superior Olivary Nucleus (SON) is the major source of descending GABAergic inhibition to the ipsilateral nucleus magnocellularis (NM), nucleus angularis (NA) and NL, and receives excitatory input from NA and NL. We used viral tracers to reveal projections from SON to NL, and characterized response types within SON. In NL, we isolated inhibitory synaptic contributions from extracellular field potentials through the iontophoretic application of blockers of GABA (gabazine) or glycine (strychnine). These blockers increase onset and offset responses evoked by tonal stimuli, but did not shift best ITD. The co-application of gabazine and strychnine revealed supra-linear summation of GABA and glycine. Profiles of synaptic activation revealed more prominent inhibition following stimulus offset, suggesting inputs to NL originate from both sustained and offset response types of the SON. These heterogeneous responses may represent separate SON neuronal populations. Together, our results support the role of SON inhibition in increasing the dynamic range of ITD tuning of NL neurons.
- 1 January 2026
- Published by SAGE Publications in Trends in Hearing
- Vol. 30
Open AccessInhibitory inputs to avian ITD circuits
In birds, detection of sound source azimuth begins in the nucleus laminaris (NL), which computes interaural time differences (ITDs). Inhibition has been proposed to protect NL neurons from losing ITD sensitivity with increasing sound level, although the nature of this inhibition is incompletely understood. Anatomical studies show GABAergic synapses throughout NL and first order nuclei, while in vitro studies in chicken NL showed that GABAergic inhibitory postsynaptic potentials (IPSPs) shorten membrane time constants, perhaps to allow the membrane potential to follow rapid synaptic currents accurately over a range of sound levels. Given the importance of inhibition in regulating auditory brainstem activity, we examined the nature of inhibitory input to NL in vivo. The Superior Olivary Nucleus (SON) is the major source of descending GABAergic inhibition to the ipsilateral nucleus magnocellularis (NM), nucleus angularis (NA) and NL, and receives excitatory input from NA and NL. We used viral tracers to reveal projections from SON to NL, and characterized response types within SON. In NL, we isolated inhibitory synaptic contributions from extracellular field potentials through the iontophoretic application of blockers of GABA (gabazine) or glycine (strychnine). These blockers increase onset and offset responses evoked by tonal stimuli, but did not shift best ITD. The co-application of gabazine and strychnine revealed supra-linear summation of GABA and glycine. Profiles of synaptic activation revealed more prominent inhibition following stimulus offset, suggesting inputs to NL originate from both sustained and offset response types of the SON. These heterogeneous responses may represent separate SON neuronal populations. Together, our results support the role of SON inhibition in increasing the dynamic range of ITD tuning of NL neurons.
- Open Access
Inhibitory inputs to avian ITD circuits
In birds, detection of sound source azimuth begins in the nucleus laminaris (NL), which computes interaural time differences (ITDs). Inhibition has been proposed to protect NL neurons from losing ITD sensitivity with increasing sound level, although the nature of this inhibition is incompletely understood. Anatomical studies show GABAergic synapses throughout NL and the first order nuclei, while in vitro studies in chicken NL showed that depolarizing, GABAergic inhibitory postsynaptic potentials (IPSPs) shorten membrane time constants, perhaps to allow the membrane potential to follow rapid synaptic currents accurately over a range of sound levels. Given the importance of inhibition in regulating auditory brainstem activity, we examined the nature of the inhibitory input to NL in vivo . The Superior Olivary Nucleus (SON) is the major source of descending GABAergic inhibition to the ipsilateral nucleus magnocellularis (NM), nucleus angularis (NA) and NL, and receives excitatory input from NA and NL. We used viral tracers to reveal projections from SON to NL, and characterized response types within SON. In NL, we isolated inhibitory synaptic contributions from extracellular field potentials through the iontophoretic application of blockers of GABA (gabazine) or glycine (strychnine). These blockers increase the onset and offset responses evoked by tonal stimuli, but did not shift best ITD. The co-application of gabazine and strychnine revealed supra-linear summation of GABA and glycine. Profiles of synaptic activation revealed more prominent inhibition following stimulus offset, suggesting inputs to NL originate from both the sustained and offset response types of the SON. These heterogeneous responses may represent separate SON neuronal populations.
Open AccessInhibitory inputs to avian ITD circuits
In birds, detection of sound source azimuth begins in the nucleus laminaris (NL), which computes interaural time differences (ITDs). Inhibition has been proposed to protect NL neurons from losing ITD sensitivity with increasing sound level, although the nature of this inhibition is incompletely understood. Anatomical studies show GABAergic synapses throughout NL and the first order nuclei, while in vitro studies in chicken NL showed that depolarizing, GABAergic inhibitory postsynaptic potentials (IPSPs) shorten membrane time constants, perhaps to allow the membrane potential to follow rapid synaptic currents accurately over a range of sound levels. Given the importance of inhibition in regulating auditory brainstem activity, we examined the nature of the inhibitory input to NL in vivo . The Superior Olivary Nucleus (SON) is the major source of descending GABAergic inhibition to the ipsilateral nucleus magnocellularis (NM), nucleus angularis (NA) and NL, and receives excitatory input from NA and NL. We used viral tracers to reveal projections from SON to NL, and characterized response types within SON. In NL, we isolated inhibitory synaptic contributions from extracellular field potentials through the iontophoretic application of blockers of GABA (gabazine) or glycine (strychnine). These blockers increase the onset and offset responses evoked by tonal stimuli, but did not shift best ITD. The co-application of gabazine and strychnine revealed supra-linear summation of GABA and glycine. Profiles of synaptic activation revealed more prominent inhibition following stimulus offset, suggesting inputs to NL originate from both the sustained and offset response types of the SON. These heterogeneous responses may represent separate SON neuronal populations.
- Open Access
- 20 August 2025
- Published by Springer Nature in Nature
- Vol. 646 (8085) , 1-9
Network synchrony creates neural filters promoting quiescence in Drosophila
Animals require undisturbed periods of rest during which they undergo recuperative processes1. However, it is unclear how brain states arise that are able to dissociate an animal from its external world, enabling quiescent behaviours, while retaining vigilance to salient sensory cues2. Here we describe a neural mechanism in Drosophila that creates neural filters that engender a brain state that enables quiescent behaviour by generating coherent slow-wave activity (SWA)3 between sleep-need4 (R5)- and locomotion-promoting neural networks5. The coherence of SWA is subject to circadian and homeostatic control and can be modulated by sensory experience. Mimicry of coherent SWA reveals that R5 oscillations reduce responsiveness to visual stimuli by rhythmically associating neural activity of locomotion-promoting cells, effectively overruling their output. These networks can regulate behavioural responsiveness by providing antagonistic inputs to downstream head-direction cells6,7. Thus, coherent oscillations provide the mechanistic basis for a neural filter by temporally associating opposing signals, resulting in reduced functional connectivity between locomotion-gating and navigational networks. We propose that the temporal pattern of SWA provides the structure to create a ‘breakable’ filter, permitting the animal to enter a quiescent state, while providing the architecture for strong or salient stimuli to ‘break’ the neural interaction, consequently allowing the animal to react.
- 20 August 2025
- Published by Springer Nature in Nature
- Vol. 646 (8085) , 1-9
Open AccessNetwork synchrony creates neural filters promoting quiescence in Drosophila
Animals require undisturbed periods of rest during which they undergo recuperative processes1. However, it is unclear how brain states arise that are able to dissociate an animal from its external world, enabling quiescent behaviours, while retaining vigilance to salient sensory cues2. Here we describe a neural mechanism in Drosophila that creates neural filters that engender a brain state that enables quiescent behaviour by generating coherent slow-wave activity (SWA)3 between sleep-need4 (R5)- and locomotion-promoting neural networks5. The coherence of SWA is subject to circadian and homeostatic control and can be modulated by sensory experience. Mimicry of coherent SWA reveals that R5 oscillations reduce responsiveness to visual stimuli by rhythmically associating neural activity of locomotion-promoting cells, effectively overruling their output. These networks can regulate behavioural responsiveness by providing antagonistic inputs to downstream head-direction cells6,7. Thus, coherent oscillations provide the mechanistic basis for a neural filter by temporally associating opposing signals, resulting in reduced functional connectivity between locomotion-gating and navigational networks. We propose that the temporal pattern of SWA provides the structure to create a ‘breakable’ filter, permitting the animal to enter a quiescent state, while providing the architecture for strong or salient stimuli to ‘break’ the neural interaction, consequently allowing the animal to react.
- Open Access
- 15 August 2025
- Published by eLife Sciences Publications, Ltd in eLife
- Vol. 13
Sub-type specific connectivity between CA3 pyramidal neurons may underlie their sequential activation during sharp waves
The CA3 region of the hippocampus is the major site of sharp wave initiation, and a brain region crucially involved in learning and memory. Highly recurrent connectivity within its excitatory network is thought to underlie processes involved in memory formation. Recent work has indicated that distinct subpopulations of pyramidal neurons within this region may contribute differently to network activity, including sharp waves, in CA3. Exactly how these contributions may arise is not yet known. Here, we disentangle the local connectivity between two distinct CA3 cell types in mice: thorny and athorny pyramidal cells. We find an asymmetry in the connectivity between these two populations, with athorny cells receiving strong input from both athorny and thorny cells. Conversely, the thorny cell population receives very little input from the athorny population. Computational modeling suggests that this connectivity scheme may determine the sequential activation of these cell types during large network events such as sharp waves.
- 15 August 2025
- Published by eLife Sciences Publications, Ltd in eLife
- Vol. 13
Open AccessSub-type specific connectivity between CA3 pyramidal neurons may underlie their sequential activation during sharp waves
The CA3 region of the hippocampus is the major site of sharp wave initiation, and a brain region crucially involved in learning and memory. Highly recurrent connectivity within its excitatory network is thought to underlie processes involved in memory formation. Recent work has indicated that distinct subpopulations of pyramidal neurons within this region may contribute differently to network activity, including sharp waves, in CA3. Exactly how these contributions may arise is not yet known. Here, we disentangle the local connectivity between two distinct CA3 cell types in mice: thorny and athorny pyramidal cells. We find an asymmetry in the connectivity between these two populations, with athorny cells receiving strong input from both athorny and thorny cells. Conversely, the thorny cell population receives very little input from the athorny population. Computational modeling suggests that this connectivity scheme may determine the sequential activation of these cell types during large network events such as sharp waves.
- Open Access
Conditions for replay of neuronal assemblies
From cortical synfire chains to hippocampal replay, the idea that neural populations can be activated sequentially with precise spike timing is thought to be essential for several brain functions. It has been shown that neuronal sequences with weak feedforward connectivity can be replayed due to amplification via intra-assembly recurrent connections. However, the mechanisms behind this phenomenon are still unclear. Here, we simulate spiking networks with different excitatory and inhibitory connectivity and find that an exclusively excitatory network is sufficient for this amplification to occur. To explain the spiking network behavior, we introduce a population model of membrane-potential distributions, and we analytically describe how different connectivity structures determine replay speed, with weaker feedforward connectivity generating slower and wider pulses that can be sustained by recurrent connections. Pulse propagation is facilitated if the neuronal membrane time constant is large compared to the pulse width. Together, our simulations and analytical results predict the conditions for replay of neuronal assemblies. Author summary: In this work, we study how neural activity can propagate through a network of neurons. We therefore consider the activities of defined groups of neurons, which are also called assemblies. We are particularly interested in understanding how a previously learned sequence of activity patterns of assemblies can be replayed. To this aim, we combine large-scale numerical simulations with simpler analytical descriptions of neural dynamics. Our simulations show that, if feedforward connections across assemblies are weak, pulses of activity can be amplified by intra-assembly recurrent connections, allowing for sequence retrieval across a wide range of network structures and parameters. We introduce a new theoretical framework to study sequential activity, deriving conditions for sequence replay to succeed, and unveiling how replay speed depends on different network parameters. Crucially, we find that subthreshold membrane potential distributions are essential in determining the properties of the activity pulse and whether replay can succeed. Our findings contribute to understanding the mechanisms of hippocampal replay in particular, which is important for memory consolidation, as well as the propagation of activity across feedforward neuronal circuits in general.
Open AccessConditions for replay of neuronal assemblies
From cortical synfire chains to hippocampal replay, the idea that neural populations can be activated sequentially with precise spike timing is thought to be essential for several brain functions. It has been shown that neuronal sequences with weak feedforward connectivity can be replayed due to amplification via intra-assembly recurrent connections. However, the mechanisms behind this phenomenon are still unclear. Here, we simulate spiking networks with different excitatory and inhibitory connectivity and find that an exclusively excitatory network is sufficient for this amplification to occur. To explain the spiking network behavior, we introduce a population model of membrane-potential distributions, and we analytically describe how different connectivity structures determine replay speed, with weaker feedforward connectivity generating slower and wider pulses that can be sustained by recurrent connections. Pulse propagation is facilitated if the neuronal membrane time constant is large compared to the pulse width. Together, our simulations and analytical results predict the conditions for replay of neuronal assemblies. Author summary: In this work, we study how neural activity can propagate through a network of neurons. We therefore consider the activities of defined groups of neurons, which are also called assemblies. We are particularly interested in understanding how a previously learned sequence of activity patterns of assemblies can be replayed. To this aim, we combine large-scale numerical simulations with simpler analytical descriptions of neural dynamics. Our simulations show that, if feedforward connections across assemblies are weak, pulses of activity can be amplified by intra-assembly recurrent connections, allowing for sequence retrieval across a wide range of network structures and parameters. We introduce a new theoretical framework to study sequential activity, deriving conditions for sequence replay to succeed, and unveiling how replay speed depends on different network parameters. Crucially, we find that subthreshold membrane potential distributions are essential in determining the properties of the activity pulse and whether replay can succeed. Our findings contribute to understanding the mechanisms of hippocampal replay in particular, which is important for memory consolidation, as well as the propagation of activity across feedforward neuronal circuits in general.
- Open Access
Population sparseness determines strength of Hebbian plasticity for maximal memory lifetime in associative networks
The brain can efficiently learn and form memories based on limited exposure to stimuli. One key factor believed to support this ability is sparse coding, which can reduce overlap between representations and minimize interference. It is well known that increased sparseness can enhance memory capacity, yet its impact on the speed of learning remains poorly understood. Here we analyze the relationship between population sparseness and learning speed — specifically, how the learning speed that maximizes memory capacity depends on the sparseness of the neural code, and how this in turn affects the network’s maximal capacity. To this end, we study a feedforward network with Hebbian and homeostatic plasticity and a two-state synapse model. The network learns to associate binary input-output pattern pairs, where sparseness corresponds to a small fraction of active neurons per pattern. The learning speed is modeled as the probability of synaptic changes during learning. The results presented in this manuscript are based on both network simulations and an analytical theory that predicts expected memory capacity and optimal learning speed. For both perfect and noisy retrieval cues, we find that the optimal learning speed indeed increases with increasing pattern sparseness — an effect that is more pronounced for input sparseness than for output sparseness. Interestingly, the optimal learning speed stays the same across different network sizes if the number of active units in an input pattern is kept constant. While the capacity obtained at optimal learning speed increases monotonically with output sparseness, its dependence on input sparseness is non-monotonic. Overall, we provide the first detailed investigation of the interactions between population sparseness, learning speed, and storage capacity. Our findings propose that differences in population sparseness across brain regions may underlie observed differences in how quickly those regions adapt and learn. Author summary: The brain can efficiently learn and form memories based on limited exposure to stimuli. One key factor believed to support this ability is the way neural circuits encode and organize information about the external world. Population sparseness refers to the phenomenon in which, in many brain regions, only a small subset of neurons is active at any given time or in response to a particular stimulus. Sparse codes are believed to reduce overlap between representations and minimize interference and can thus enhance the storage capacity of a network. Here we investigate the effect of population sparseness on the speed of learning input-output associations in a network model with Hebbian learning. We find that the learning speed that yields the maximal capacity increases with increasing sparseness. The maximal capacity increases for sparser input and output representations but in the first case only up to a certain point before sparseness becomes destructive. These findings could contribute to explaining why some brain regions learn faster than others.
Open AccessPopulation sparseness determines strength of Hebbian plasticity for maximal memory lifetime in associative networks
The brain can efficiently learn and form memories based on limited exposure to stimuli. One key factor believed to support this ability is sparse coding, which can reduce overlap between representations and minimize interference. It is well known that increased sparseness can enhance memory capacity, yet its impact on the speed of learning remains poorly understood. Here we analyze the relationship between population sparseness and learning speed — specifically, how the learning speed that maximizes memory capacity depends on the sparseness of the neural code, and how this in turn affects the network’s maximal capacity. To this end, we study a feedforward network with Hebbian and homeostatic plasticity and a two-state synapse model. The network learns to associate binary input-output pattern pairs, where sparseness corresponds to a small fraction of active neurons per pattern. The learning speed is modeled as the probability of synaptic changes during learning. The results presented in this manuscript are based on both network simulations and an analytical theory that predicts expected memory capacity and optimal learning speed. For both perfect and noisy retrieval cues, we find that the optimal learning speed indeed increases with increasing pattern sparseness — an effect that is more pronounced for input sparseness than for output sparseness. Interestingly, the optimal learning speed stays the same across different network sizes if the number of active units in an input pattern is kept constant. While the capacity obtained at optimal learning speed increases monotonically with output sparseness, its dependence on input sparseness is non-monotonic. Overall, we provide the first detailed investigation of the interactions between population sparseness, learning speed, and storage capacity. Our findings propose that differences in population sparseness across brain regions may underlie observed differences in how quickly those regions adapt and learn. Author summary: The brain can efficiently learn and form memories based on limited exposure to stimuli. One key factor believed to support this ability is the way neural circuits encode and organize information about the external world. Population sparseness refers to the phenomenon in which, in many brain regions, only a small subset of neurons is active at any given time or in response to a particular stimulus. Sparse codes are believed to reduce overlap between representations and minimize interference and can thus enhance the storage capacity of a network. Here we investigate the effect of population sparseness on the speed of learning input-output associations in a network model with Hebbian learning. We find that the learning speed that yields the maximal capacity increases with increasing sparseness. The maximal capacity increases for sparser input and output representations but in the first case only up to a certain point before sparseness becomes destructive. These findings could contribute to explaining why some brain regions learn faster than others.