Background

Fig.1Spike travels down an axon, and is converted via low-pass filter to a current pulse that charges a capacitor much like a postsynaptic potential in a real neuron.8

With an explosion of interest and research in recent decades, the field of Neuroscience would benefit from a ubiquitous standard in network modeling to investigate neural properties at large scales. The most physiologically representative model to date is the pulse-coded spiking neural network described above in Figure 1. Despite the biological realism, building these networks can be difficult for neurophysiologists and researchers without prior network engineering or computer science and engineering experience. The result is poorly documented and sparsely commented simulation attempts that lack a foundation for future hardware implementation. This directly impacts the reproducibility of research, both by the reviewers, as well as the publishers themselves.

Objectives

Create a Neural Network Simulator that:

1.Models a pulse-coded network

2.Incorporates input from interviewees

3.Is easy to use and intuitive to learn

4.Produces a variety of data representations

5.Allows for dynamic network visualization

6.Offers a variety of connectivity settings

7.Affords ubiquity as a standard by using opensource software packages

Measures of success were delivered through the verification and reproduction of published research, as well as application to current research

DETAILS

Design goal and constraints

Critical Functionalities

The ultimate design has been conceptualized to have, inherent in itself, certain characteristics such

as minimal cost, high feasibility, and ease of access. These can be achieved through careful selection

of our materials and software package. The design will have components that are powerful, cheap,

and ubiquitous such that anyone wanting to create this design can do so. This design will utilize

open-source software packages specifically designed for neural network modeling, and individual

laptops to achieve the low costs required by the design considerations. The goal is to make the

design so feasible as to encourage a standard in neural network modeling. The design must also

successfully propagate an input signal or function without significant loss through several layers of

neurons. Signal preservation can be accomplished by imposing a small amount of noise on the

signal, the enhancement of input stimulus level in the first layer, as well as the conductance change

across the layers. Furthermore, the design should be able to accomplish complex tasks and present

meaningful pathway/simulation data and the proper visualization showing network structures and

graphs.

Other Functionalities

Other functionalities include the ability to customize the network model generation depending on the

user’s specific research interests. Specific input stimulus and background noisy current level, the

order of filtered input signal, axon resistances, time of running, the conductance change over the

layer, and number of neurons or layers are some examples of functions that should be specified by

the user to better represent their research subject.

Constraints

Design constraints come in the form of access to expert guidance. Many of the topics required for

this design are niche and require extensive review of the literature, such as the implementation and

instruction of the python package. In addition, program functionality is constrained by the functions

provided by the BRIAN toolbox, tkinter package, and python language. More specifically, since some

parameters of our model is hard coded, It is hard for GUI to include some function that can modify

the hard coded part. These functions are also constrained by the CPU and GPU performance of

users’ computer. Even if the end-user might need something specific, if there isn’t a class for it, or if

the computer doesn’t have the capacity to execute it, it cannot be done within the scope of this

design.

Diwei Xiong is the editor of this page