P-113 Biologically plausible functional spiking networks with efficient coding

Background: Understanding how computations performed by neural
microcircuits emerge from neural dynamics is a fundamental
question in neuroscience. Recently, the framework of efficient coding
proposed a theory of how spiking neural networks can maximize
information encoding while keeping spike numbers low to minimize
metabolic expenditure. However, while recent developments (Liu
and Pehlevan, Adv. Neural Inf. Process. Syst. 2020) incorporated elements
of biological realism such as including excitatory (E) and inhibitory
(I) spiking neurons, these models did not include key single
neuron and network properties of real biological circuits such as
adaptation currents, hyperpolarization-triggered rebound currents
and fast and slow synaptic currents. The lack of local currents such
as the adaptation and hyperpolarization-triggered currents is consequential
to the fact that their model does not take into account the
metabolic expenditure incurred by spiking.
Objective: Our objective is to develop spiking neural network models
with efficient coding with higher levels of biological realism that
predict key structural and dynamical features of neural microcircuits
dedicated to processing of sensory signals.
Methods: Unlike previous approaches of efficient coding spike
spikes (Boerlin et al., PLoS Comp. Bio. 2013; Koren et al., PLoS
Comp. Bio. 2017), we here use two objective functions that relate
to the activity of E and I neurons, respectively, and describe network’s
computation. From these, we analytically develop mathematically
closed and complete set of dynamical equations for the
membrane potentials and firing thresholds. Our results are verified
with simulations, using biologically plausible network parameters.
Results: We find a biologically plausible spiking model realizing efficient
coding in the case of a generalized leaky integrate-and-fire (LIF)network with E and I units, equipped with heterogeneous fast and
slow synaptic currents, local homeostatic currents such as spiketriggered
adaptation, hyperpolarization-activated rebound current,
heterogeneous firing thresholds and resets, heterogeneous postsynaptic
potentials, and structured, low-rank connectivity. In particular,
the connectivity is such that neurons with strongly similar
selectivity have strong synapses, neurons with weakly similar selectivity
have weak synapses and neurons with opposite selectivity are
unconnected. We show how the complexity (rank) of E-E connectivity
shapes network responses. With low rank of E-E connectivity
matrix, only stimulus-selective neurons (driven by the feedforward
current) respond to the external stimuli. With higher rank of these
interactions, also non-selective neurons respond. In the latter case,
the connectivity implements linear mixing of stimulus features in
patterns of spike trains across the population.
Conclusions: Using efficient coding theory and enforcing a biologically
plausible E-I architecture, we found a generalized LIF model
with structured connectivity. The network generates realistic and
highly non-linear responses to the stimulus. Such a biologically plausible
spiking network can be used to model neural dynamics and computations in healthy and disturbed biological microcircuits.