Biological neural networks are the basic foundation to the great computational ability of the brain.
Every biological unit contains neuron & synapses nets.
The assumption is that if we will succeed in reflecting precisely the neural network in a computational model, we will be able to understand the entire brain behavior and to reconstruct its abilities.
This project come to explore the neuron itself, and its behaviour as part of a neuron network.
Abstract
Biological neural networks are the basic foundation to the great computational ability of the brain.
Every biological unit contains neuron & synapses nets.
The assumption is that if we will succeed in reflecting precisely the neural network in a computational model, we will be able to understand the entire brain behavior and to reconstruct its abilities.
This project come to explore the neuron itself, and its behaviour as part of a neuron network.
In this project, we reffer into two different neuron, well known, computational models. The first is the LIF (leaky integrated fire) model, and the second is the Izhikevich model.
The LIF model is a simple model, based on the the electric model of capacitor and resistor as shown below.
The Izhikevich model is a pure mathematical model, which was developed by Eugene M. Izhikevich.
We focus on investigating the two different mechanism, Selective Adaptation vs Frequency Adaptation, that were described by Pr. Shimon Marom in his article.
This project is a continuation of the project “”biological realistic neural network simulation”” by Evgeny Litvin, Spring/04, and we rely on its conclusions and outcomes.
The problem (or the background)
We try to reproduce and investigate the phenomena: Frequency adaptation (FA) and Selective adaptation (SA) that were found in experiments and are described in the article “”Selective Adaptation in Networks of Cortical Neurons”” by Danny Eytan, Naama Brenner and Shimon Marom.
Frequency Adaptation
FA is a phenomenon which occurs in general neural networks.
It occurs when electrodes are attached to the neural network, and current pulses are applied in a constant frequency.
The network response is measured and calculated.
The longest the input current persists, the more the network adapts to the input at that particular frequency, and response less to the input.
Selective Adaptation
SA is a phenomenon which also occurs in general neural networks.
2 electrodes are attached to the neural network, at two far points.
The First electrode drives high frequency (0.2 Hz) pulses, while the second electrode drives low frequency (0.02 Hz).
The intensity of the response was measured.
The graph below demonstrate the initial response vs the final response for each frequency.
It is shown in the graph that the network response to high frequency weakens as times pass, whereas the response to the low frequency strengthens.
The solution (or the basic approach)
- Understanding the Spnet++ simulator.
- Reproducing Frequency Adaptation phenomena on Izhikevich neurons, using the Spnet++ simulator
- Reproducing Selective Adaptation phenomena on Izhikevich neurons, using the Spnet++ simulator
- Understanding the csim simulator
- Reproducing Frequency Adaptation phenomena on LIF neurons, using the csim simulator
- Reproducing Selective Adaptation phenomena on LIF neurons, using the csim simulator
- Determine the conditions to achieve FA and SA
- Suggest an explanation to the SA mechanism
- Determine whether there is or there isn’t any relevancy between the Selective Adaptation and the Frequency Adaptation
Tools
In this project we have used two different simulators, both based on Matlab & C++.
The first is Spnet++ simulator which was built by Evgeny, and can be downloaded from the site: http://visl.technion.ac.il/projects/2004s20/. The second is the csim /circuit simulator which can be downloaded from http://www.lsm.tugraz.at/download/index.html.
The Matlab was used also for analyzing the experiments data, & for plotting the results.
Simulations results
Below, Frequency Adaptation result, using the Csim simulator.
The simulation last 600 seconds
Network size is 1k, 80% exc neurons, 20% inh neurons
Stimulus was connected to the excitatory pool only.
parameters:
inh pool to exc pool: Cscale=1, lambda=inf,
stimulus to exc pool: rscale=0, wscale=2.5, cscale=1, lambda=inf,
Below, Selective Adaptation results, using the Csim simulator
The simulation last 600 seconds
frequncies are 20Hz and 0.1Hz
We can see SA in the LIF neurons, but not in the IZH neurons
The simulation last 600 seconds
frequncies are 20Hz and 0.1Hz
We can see SA in the IZH neurons, but not in the LIF neurons
Conclusions
- The SA is not dependant upon the existence of an electrical synapse.
we have observed SA also in a pool without electric synapse.
One of the assumption of S. Marom was that SA can be achieved only in this synapse. - The SA can be achieved in the simplest neural model – LIF, and there is no need to use a more complicated neural model as Izhikevich.
while Marom thought that SA cannot be acieved in LIF neuron, cause its a simple computed model - As is in Maroms article, SA is dependant upon FA. There is no SA without FA
- Different simulators give different results for the same neural model (Izhikevich in Spnet and in csim for FA)
- FA exist through different mechanisms at different frequencies.
We have simulated variouse of frequencies, start from 1/100 to 50.
In all of those frequencies, a FA was predicted. - When neurons exhibit adaptation for a continuous input, they will not necessarily exhibit the same behavior for repetitive pulses at a constant frequency
Acknowledgment
We are grateful to our project supervisors Karina Odinaev & Igal Raichelgauz for their help and guidance throughout this work, and
We would like to thanks Evgeny and Alik for their explanations during the project, and
We are also grateful to the personnel of the VISL lab for the great support and help.