Excitation-Inhibition Balanced Neural Networks for Fast Signal Detection

Excitation-Inhibition Balance, Neural Criticality, and Activities in Neuronal Circuits

Neural activities in local circuits exhibit complex and multilevel dynamic features. Individual neurons spike irregularly, which is believed to originate from receiving balanced amounts of excitatory and inhibitory inputs, known as the excitation-inhibition balance. The spatial-temporal cascades of clustered neuronal spikes occur in variable sizes and durations, manifested as neural avalanches with scale-free features. These may be explained by the neural criticality hypothesis, which posits that neural systems operate around the transition between distinct dynamic states. Here, we summarize the experimental evidence for and the underlying theory of excitation-inhibition balance and neural criticality. Furthermore, we review recent studies of excitatory-inhibitory networks with synaptic kinetics as a simple solution to reconcile these two apparently distinct theories in a single circuit model. This provides a more unified understanding of multilevel neural activities in local circuits, from spontaneous to stimulus-response dynamics.

A disinhibitory circuit mechanism explains a general principle of peak performance during mid-level arousal

Perceptual decision-making is highly dependent on the momentary arousal state of the brain, which fluctuates over time on a scale of hours, minutes, and even seconds. The textbook relationship between momentary arousal and task performance is captured by an inverted U-shape, as put forward in the Yerkes-Dodson law. This law suggests optimal performance at moderate levels of arousal and impaired performance at low or high arousal levels. However, despite its popularity, the evidence for this relationship in humans is mixed at best. Here, we use pupil-indexed arousal and performance data from various perceptual decision-making tasks to provide converging evidence for the inverted U-shaped relationship between spontaneous arousal fluctuations and performance across different decision types (discrimination, detection) and sensory modalities (visual, auditory). To further understand this relationship, we built a neurobiologically plausible mechanistic model and show that it is possible to reproduce our findings by incorporating two types of interneurons that are both modulated by an arousal signal. The model architecture produces two dynamical regimes under the influence of arousal: one regime in which performance increases with arousal and another regime in which performance decreases with arousal, together forming an inverted U-shaped arousal-performance relationship. We conclude that the inverted U-shaped arousal-performance relationship is a general and robust property of sensory processing. It might be brought about by the influence of arousal on two types of interneurons that together act as a disinhibitory pathway for the neural populations that encode the available sensory evidence used for the decision.

Distinctive properties of biological neural networks and recent advances in bottom-up approaches toward a better biologically plausible neural network

Although it may appear infeasible and impractical, building artificial intelligence (AI) using a bottom-up approach based on the understanding of neuroscience is straightforward. The lack of a generalized governing principle for biological neural networks (BNNs) forces us to address this problem by converting piecemeal information on the diverse features of neurons, synapses, and neural circuits into AI. In this review, we described recent attempts to build a biologically plausible neural network by following neuroscientifically similar strategies of neural network optimization or by implanting the outcome of the optimization, such as the properties of single computational units and the characteristics of the network architecture. In addition, we proposed a formalism of the relationship between the set of objectives that neural networks attempt to achieve, and neural network classes categorized by how closely their architectural features resemble those of BNN. This formalism is expected to define the potential roles of top-down and bottom-up approaches for building a biologically plausible neural network and offer a map helping the navigation of the gap between neuroscience and AI engineering.

Dorsolateral Prefrontal Cortex Glutamate/Gamma-Aminobutyric Acid (GABA) Alterations in Clinical High Risk and First-Episode Schizophrenia: A Preliminary 7-T Magnetic Resonance Spectroscopy Imaging Study

Converging lines of evidence suggest that an imbalance between excitation and inhibition is present in the dorsolateral prefrontal cortex (DLPFC) of schizophrenia (SCZ). Gamma-aminobutyric-acid (GABA) and, to a lesser extent, glutamate (Glu) abnormalities were reported in the DLPFC of SCZ patients, especially on the right hemisphere, by post-mortem studies. However, in vivo evidence of GABA, Glu, and Glu/GABA DLPFC abnormalities, particularly on the right side and the early stages of illness, is limited. In this preliminary study, we utilized 7-Tesla magnetic resonance spectroscopic imaging (MRSI) to investigate bilateral Glu/Creatine (Cre), GABA/Cre, and Glu/GABA in the DLPFC of sixteen first episode schizophrenia (FES), seventeen clinical high risk (CHR), and twenty-six healthy comparison (HC) subjects. FES and CHR had abnormal GABA/Cre and Glu/GABA in the right DLPFC (rDLPFC) compared with HC participants, while no differences were observed in the left DLPFC (lDLPFC) among the three groups. Furthermore, HC had higher Glu/GABA in rDLPFC compared to lDLPFC (R > L), whereas the opposite relationship (R < L) was observed in the DLPFC Glu/GABA of FES patients. Altogether, these findings indicate that GABA/Cre and Glu/GABA DLPFC alterations are present before illness manifestation and worsen in FES patients, thus representing a putative early pathophysiological biomarker for SCZ and related psychotic disorders.

On the Role of the Excitation/Inhibition Balance of Homeostatic Artificial Neural Networks

Homeostatic models of artificial neural networks have been developed to explain the self-organization of a stable dynamical connectivity between the neurons of the net. These models are typically two-population models, with excitatory and inhibitory cells. In these models, connectivity is a means to regulate cell activity, and in consequence, intracellular calcium levels towards a desired target level. The excitation/inhibition (E/I) balance is usually set to 80:20, a value characteristic for cortical cell distributions. We study the behavior of these homeostatic models outside of the physiological range of the E/I balance, and we find a pronounced bifurcation at about the physiological value of this balance. Lower inhibition values lead to sparsely connected networks. At a certain threshold value, the neurons develop a reasonably connected network that can fulfill the homeostasis criteria in a stable way. Beyond the threshold, the behavior of the artificial neural network changes drastically, with failing homeostasis and in consequence with an exploding number of connections. While the exact value of the balance at the bifurcation point is subject to the parameters of the model, the existence of this bifurcation might explain the stability of a certain E/I balance across a wide range of biological neural networks. Assuming that this class of models describes the self-organization of biological network connectivity reasonably realistically, the omnipresent physiological balance might represent a case of self-organized criticality in order to obtain a good connectivity while allowing for a stable intracellular calcium homeostasis.

Efficient and robust coding in heterogeneous recurrent networks

Cortical networks show a large heterogeneity of neuronal properties. However, traditional coding models have focused on homogeneous populations of excitatory and inhibitory neurons. Here, we analytically derive a class of recurrent networks of spiking neurons that close to optimally track a continuously varying input online, based on two assumptions: 1) every spike is decoded linearly and 2) the network aims to reduce the mean-squared error between the input and the estimate. From this we derive a class of predictive coding networks, that unifies encoding and decoding and in which we can investigate the difference between homogeneous networks and heterogeneous networks, in which each neurons represents different features and has different spike-generating properties. We find that in this framework, 'type 1' and 'type 2' neurons arise naturally and networks consisting of a heterogeneous population of different neuron types are both more efficient and more robust against correlated noise. We make two experimental predictions: 1) we predict that integrators show strong correlations with other integrators and resonators are correlated with resonators, whereas the correlations are much weaker between neurons with different coding properties and 2) that 'type 2' neurons are more coherent with the overall network activity than 'type 1' neurons.