SIMULATION OF NOISE INFLUENCE ON LEARNING PROCESS IN TWO-DIMENSIONAL BIOLOGICAL NEURAL NETWORKS

Tatsiana Pyzhyk1, Andrey Denisov1

1 Department of Physics, Belarusian State University, Belarus

An understanding of brain processes mechanisms has a great importance for formation of general ideas about a brain and for the possibility of treatment and prevention of neuropathology [1]. Learning is the one of the most important process which takes place in the mammalian brain. Just the ability to learn is the basis for survival in the ever-changing environment. Each pyramidal neuron in a brain has thousands of synaptic contacts in cortex and approximately 80% of them are excitatory and 20% are inhibitory. The balance of excitation and inhibition signals is a very important factor [2]. This phenomenon is based on neuroplasticity and is realized due to homeostatic tuning of synaptic contacts strength [3]. In addition, an important feature of a biological network is that cortical neurons in vivo are subjected to an ongoing stimulation by input synaptic activity which is the main source of a low-intensity noise and sustains membrane potentials of cortical neurons in a noisy and depolarized state [4-6]. The computer model of neural network with similar to biological parameters was developed to investigate the influence of noise and inhibition on the learning process.

The model is realized in the Neural Simulation Tool (NEST) simulation software. The model network consists of the 100×100 nodal grid of excitatory neurons and the 50×50 nodal grid of inhibitory neurons, which are superimposed on each other. Neurons are simulated on the basis of integrate-and-fire (IAF) model. Excitatory synaptic contacts are based on the spike time dependent plasticity (STDP) rule, and inhibitory contacts are implemented as stationary constant strength connections whose value equal to half of the maximum of excitatory connections strength. Initial values of excitatory connections are distributed evenly within minimal and maximal strengths. Connections in the model are of each-to-each type with constant probability. Delay of propagation depends on a distance between cells linearly. A noise component is included as high frequency Poisson generators connected to each neuron. It is similar to uniform background subliminal activity in biological networks. The resulting model has important properties of biological neural network, such as dense local recurrent connections [2]. The training signals are represented as repetitive simultaneous activations of cells groups. Two training signals are set in opposite quarters of network with different time intervals.

The effect of learning is estimated as the ratio of average membrane potential of two learning zones in an inter-stimuli period. Both zones are in opposite quarter to last presented signal. The first one includes neurons on which other stimulus was set before. The second one consists of other cells in the same quarter. Learning is observed in case when the coefficient of learning is greater than one. In this case, learning is defined as more strong response of previously trained cells under the influence of excitation waves which propagate from the place of delivery of another stimulus.

It was shown that a noise impact and/or a high intensity of the training signal could lead the network to the state with a maintained epileptic activity in case of the absence or weak inhibition. The introduction of inhibition makes it possible to observe training at those values of noise and/or signal intensity at which training was impossible earlier. It also leads to suppression of an epileptic activity if such activity arises. However, large values of the inhibiting weights strength lead to the conditions when most of the neurons are in the state with a membrane potential well below the threshold potential. It makes learning impossible. A positive effect of noise exposure is that a subthreshold activation of neurons throughout the network is caused by a wave of induced activity propagating from the first training signal, but the most significant changes in the membrane potential occur in the region corresponding to the second training signal. Over-threshold activation occurs at a certain ratio of excitation and noise in the region corresponding to the second training signal. It leads to the fixation of previously learned signals (because of STDP). Therefore, the network-learning coefficient depends on ratio of such parameters as strength of excitatory and inhibitory synaptic contacts, intensity of noise and training signal. The proposed model can be extended by introducing plasticity of inhibitory neurons.


[1] G. Neves, S. F. Cooke, T. V. P. Bliss., Synaptic plasticity, memory and the hippocampus: a neural network approach to causality, Nature Reviews Neuroscience 9, 65-75 (2008).

[2] A. Renart., J. De La Rocha, P. Bartho et al., The asynchronous state in cortical circuits, Science 327, 587-590 (2010).

[3] J. Barral, Al. D. Reyes, Synaptic scaling rule preserves excitatory-inhibitory balance and salient neuronal network dynamics, Nature neuroscience 19, 1690-1696 (2016).

[4] N. Fourcaud, N. Brunel, Dynamics of the firing probability of noisy integrate-and-fire neurons, Neural computation 14, 2057-2110 (2002).

[5] B. Haider, A. Duque, A. R. Hasenstaub et al., Neocortical network activity in vivo is generated through a dynamic balance of excitation and inhibition, Journal of Neuroscience 26, 4535-4545 (2006).

[6] N. Brunel, Dynamics of sparsely connected networks of excitatory and inhibitory spiking neurons, Journal of computational neuroscience 8, 183-208 (2000).