P-113 Biologically plausible functional spiking networks with efficient coding
Standard
P-113 Biologically plausible functional spiking networks with efficient coding. / Koren, Veronika; Panzeri, Stefano.
In: CLIN NEUROPHYSIOL, Vol. 148, 2023, p. e59-60.Research output: SCORING: Contribution to journal › Conference abstract in journal › Research › peer-review
Harvard
APA
Vancouver
Bibtex
}
RIS
TY - JOUR
T1 - P-113 Biologically plausible functional spiking networks with efficient coding
AU - Koren, Veronika
AU - Panzeri, Stefano
PY - 2023
Y1 - 2023
N2 - Background: Understanding how computations performed by neuralmicrocircuits emerge from neural dynamics is a fundamentalquestion in neuroscience. Recently, the framework of efficient codingproposed a theory of how spiking neural networks can maximizeinformation encoding while keeping spike numbers low to minimizemetabolic expenditure. However, while recent developments (Liuand Pehlevan, Adv. Neural Inf. Process. Syst. 2020) incorporated elementsof biological realism such as including excitatory (E) and inhibitory(I) spiking neurons, these models did not include key singleneuron and network properties of real biological circuits such asadaptation currents, hyperpolarization-triggered rebound currentsand fast and slow synaptic currents. The lack of local currents suchas the adaptation and hyperpolarization-triggered currents is consequentialto the fact that their model does not take into account themetabolic expenditure incurred by spiking.Objective: Our objective is to develop spiking neural network modelswith efficient coding with higher levels of biological realism thatpredict key structural and dynamical features of neural microcircuitsdedicated to processing of sensory signals.Methods: Unlike previous approaches of efficient coding spikespikes (Boerlin et al., PLoS Comp. Bio. 2013; Koren et al., PLoSComp. Bio. 2017), we here use two objective functions that relateto the activity of E and I neurons, respectively, and describe network’scomputation. From these, we analytically develop mathematicallyclosed and complete set of dynamical equations for themembrane potentials and firing thresholds. Our results are verifiedwith simulations, using biologically plausible network parameters.Results: We find a biologically plausible spiking model realizing efficientcoding in the case of a generalized leaky integrate-and-fire (LIF)network with E and I units, equipped with heterogeneous fast andslow synaptic currents, local homeostatic currents such as spiketriggeredadaptation, hyperpolarization-activated rebound current,heterogeneous firing thresholds and resets, heterogeneous postsynapticpotentials, and structured, low-rank connectivity. In particular,the connectivity is such that neurons with strongly similarselectivity have strong synapses, neurons with weakly similar selectivityhave weak synapses and neurons with opposite selectivity areunconnected. We show how the complexity (rank) of E-E connectivityshapes network responses. With low rank of E-E connectivitymatrix, only stimulus-selective neurons (driven by the feedforwardcurrent) respond to the external stimuli. With higher rank of theseinteractions, also non-selective neurons respond. In the latter case,the connectivity implements linear mixing of stimulus features inpatterns of spike trains across the population.Conclusions: Using efficient coding theory and enforcing a biologicallyplausible E-I architecture, we found a generalized LIF modelwith structured connectivity. The network generates realistic andhighly non-linear responses to the stimulus. Such a biologically plausiblespiking network can be used to model neural dynamics and computations in healthy and disturbed biological microcircuits.
AB - Background: Understanding how computations performed by neuralmicrocircuits emerge from neural dynamics is a fundamentalquestion in neuroscience. Recently, the framework of efficient codingproposed a theory of how spiking neural networks can maximizeinformation encoding while keeping spike numbers low to minimizemetabolic expenditure. However, while recent developments (Liuand Pehlevan, Adv. Neural Inf. Process. Syst. 2020) incorporated elementsof biological realism such as including excitatory (E) and inhibitory(I) spiking neurons, these models did not include key singleneuron and network properties of real biological circuits such asadaptation currents, hyperpolarization-triggered rebound currentsand fast and slow synaptic currents. The lack of local currents suchas the adaptation and hyperpolarization-triggered currents is consequentialto the fact that their model does not take into account themetabolic expenditure incurred by spiking.Objective: Our objective is to develop spiking neural network modelswith efficient coding with higher levels of biological realism thatpredict key structural and dynamical features of neural microcircuitsdedicated to processing of sensory signals.Methods: Unlike previous approaches of efficient coding spikespikes (Boerlin et al., PLoS Comp. Bio. 2013; Koren et al., PLoSComp. Bio. 2017), we here use two objective functions that relateto the activity of E and I neurons, respectively, and describe network’scomputation. From these, we analytically develop mathematicallyclosed and complete set of dynamical equations for themembrane potentials and firing thresholds. Our results are verifiedwith simulations, using biologically plausible network parameters.Results: We find a biologically plausible spiking model realizing efficientcoding in the case of a generalized leaky integrate-and-fire (LIF)network with E and I units, equipped with heterogeneous fast andslow synaptic currents, local homeostatic currents such as spiketriggeredadaptation, hyperpolarization-activated rebound current,heterogeneous firing thresholds and resets, heterogeneous postsynapticpotentials, and structured, low-rank connectivity. In particular,the connectivity is such that neurons with strongly similarselectivity have strong synapses, neurons with weakly similar selectivityhave weak synapses and neurons with opposite selectivity areunconnected. We show how the complexity (rank) of E-E connectivityshapes network responses. With low rank of E-E connectivitymatrix, only stimulus-selective neurons (driven by the feedforwardcurrent) respond to the external stimuli. With higher rank of theseinteractions, also non-selective neurons respond. In the latter case,the connectivity implements linear mixing of stimulus features inpatterns of spike trains across the population.Conclusions: Using efficient coding theory and enforcing a biologicallyplausible E-I architecture, we found a generalized LIF modelwith structured connectivity. The network generates realistic andhighly non-linear responses to the stimulus. Such a biologically plausiblespiking network can be used to model neural dynamics and computations in healthy and disturbed biological microcircuits.
U2 - 10.1016/j.clinph.2023.02.130
DO - 10.1016/j.clinph.2023.02.130
M3 - Conference abstract in journal
VL - 148
SP - e59-60
JO - CLIN NEUROPHYSIOL
JF - CLIN NEUROPHYSIOL
SN - 1388-2457
ER -