KR20230089495A - Spiking neural network input data processing system in GUI-based integrated development environment - Google Patents

Abstract

The present invention relates to a system for processing input data of a spiking neural network, and more particularly, in a neuromorphic-based development environment such as NADIE, an autonomous IoT integrated development environment, by simply inputting necessary parameter values by a user, an encoder and a decoder. It relates to a spiking neural network input data processing system of a GUI-based integrated development environment that provides a GUI environment in which components can be used.

Description

Spiking neural network input data processing system in GUI-based integrated development environment}

The present invention relates to a system for processing input data of a spiking neural network, and more particularly, in a neuromorphic-based development environment such as NADIE, an autonomous IoT integrated development environment, by simply inputting necessary parameter values by a user, an encoder and a decoder. It relates to a spiking neural network input data processing system of a GUI-based integrated development environment that provides a GUI environment in which components can be used.

Recently, with the development of the 4th industrial revolution, the use of artificial neural networks in the Internet of Things environment is increasing. However, in the case of most artificial neural network models in the IoT environment, power consumption is limited. A Deep Neural Network (DNN) has at least two neuron layers, and a Convolution Neural Network (CNN) model may have 64 layers. The neuron-synaptic array type alone increases hardware complexity, and this increase in hardware complexity also increases power consumption.

On the other hand, Spiking Neural Network (SNN) mimics the human brain computational process and applies it to synapses in the form of pulses. It does not require multi-layer neurons, so hardware complexity can be minimized and power consumption can be greatly reduced. . Artificial Neural Network (ANN) models, including DNN models, have been used as second-generation artificial neural networks, and spiking neural networks are attracting attention as next-generation artificial neural networks in the resource-limited IoT environment.

In order to provide services that combine the IoT environment and artificial neural networks, it is essential to support IoT application development based on neuromorphic architecture on one platform. Most platforms that support IoT application development provide visualization tools such as data analysis and statistics, but the usability of machine learning is lacking. Simulation tools such as Nengo are focused on modeling brain structures and synapses and analyzing behaviors in the brain science area, so there are limitations in applying machine learning to the IoT area. Therefore, NAIDE, an autonomous IoT integrated development environment, was developed to support neuromorphic hardware in the IoT platform and develop IoT applications.

NAIDE is a neuromorphic-based autonomous IoT application integrated development environment, and implements a program by configuring a flow by arranging pre-implemented components and controlling data flow. In addition, it supports the implementation of artificial neural network models by deploying intelligent components.

Such NAIDE has been proposed as Korean Patent Registration No. 10-2188044. Conventional technology requires a process of converting static input data of an existing artificial neural network into dynamic data containing time-series information in order to provide a spiking neural network model, and NAIDE's application development accompanying this data conversion process is The level of difficulty was high.

Korean Patent Registration No. 10-2188044 (2020.12.01.)

The present invention, which was devised to solve such a problem, relates to an input data processing system for a spiking neural network, and more specifically, parameters required by users in a neuromorphic-based development environment such as NADIE, an autonomous IoT integrated development environment. It relates to a spiking neural network input data processing system of a GUI-based integrated development environment that provides a GUI environment in which encoder and decoder components can be used simply by inputting values.

Spiking neural network input data processing system of GUI-based integrated development environment according to an embodiment of the present invention for achieving the above object,

an input device 110 that receives component information;

a processor 120 processing component information input from the input device; and

A memory device 130 for storing component information processed by the processor 120,

The processor 120 is composed of an encoder component 121 and a decoder component 122,

It is characterized in that the encoder component 121 provides a static-dynamic data conversion function.

At this time, the input device 110 may provide a graphic user interface (GUI).

Also, the encoder component 121 may be configured to encode 2D static data into dynamic data including a time interval based on parameters input from the input device 110 .

In addition, the encoder component 121 may be configured to use one or more encoding methods selected from a simple encoding method, a Poisson encoding method, and a population encoding method.

Additionally, the decoder component 122 may be configured to reconvert encoded dynamic data to static data.

In the spiking neural network input data processing system of the GUI-based integrated development environment of the present invention, the input device supports a graphic user interface to input necessary parameters without separate code input, enabling application development to which a neuromorphic architecture is applied. Therefore, in order to develop an application to which a neuromorphic architecture is applied, less time is required compared to the time required by a developer to directly code, thereby maximizing development efficiency. In addition, it has the advantage of lowering the difficulty of development because it is possible to perform application development even with basic knowledge outside of the conventional development environment in which a skilled developer is required for application development.

The user can use the encoder and decoder components just by inputting the necessary parameter values, and the program can be easily implemented by simply arranging the components without prior learning by the user in the GUI environment. In addition, both application development based on static data and dynamic data can be supported through the encoder component and the decoder component.

In addition, three types of encoding components are provided so that they can be used as needed to maximize utilization, and the encoded data through the encoder can be provided in at least one of the three ways according to the spiking neural network model, improving compatibility. can be maximized.

In addition, as a result of performing decoding using the parameter values used in the encoding process of the decoder component, static data with a high similarity to the original data can be obtained compared to the result without using the parameter values, thereby identifying and determining the accuracy of the generated dynamic data. this is possible

1 is an embodiment of a spiking neural network input data processing system of the GUI-based integrated development environment of the present invention.
Figure 2 is an embodiment of the input device of the present invention
3 is an embodiment of a simple encoding scheme of the present invention
4 is an embodiment of the Poisson encoding method of the present invention
5 is an embodiment of the population encoding method of the present invention
6 is one embodiment of a decoder component of the present invention;

Hereinafter, the present invention will be described in more detail with reference to the accompanying drawings. Prior to this, the terms or words used in this specification and claims should not be construed as being limited to ordinary or dictionary meanings, and the inventor appropriately uses the concept of terms in order to explain his/her invention in the best way. Based on the principle that it can be defined, it should be interpreted as meaning and concept consistent with the technical spirit of the present invention. In addition, unless there is another definition in the technical terms and scientific terms used, they have meanings commonly understood by those of ordinary skill in the art to which this invention belongs, and the gist of the present invention is described in the following description and accompanying drawings. Descriptions of well-known functions and configurations that may be unnecessarily obscure are omitted. The drawings introduced below are provided as examples to sufficiently convey the spirit of the present invention to those skilled in the art. Accordingly, the present invention may be embodied in other forms without being limited to the drawings presented below. Also, like reference numerals denote like elements throughout the specification. It should be noted that like elements in the drawings are indicated by like numerals wherever possible.

Figure 1 is an embodiment of the spiking neural network input data processing system of the GUI-based integrated development environment of the present invention, Figure 2 is an example of the input device of the present invention, Figure 3 is one of the simple encoding method of the present invention 4 is an embodiment of the Poisson encoding method of the present invention, FIG. 5 is an embodiment of the Population encoding method of the present invention, and FIG. 6 is an embodiment of the decoder component of the present invention.

The spiking neural network input data processing system of the GUI-based integrated development environment of the present invention,

an input device 110 that receives component information;

a processor 120 processing component information input from the input device 110; and

A memory device 130 for storing component information processed by the processor 120 may be included.

As shown in FIG. 1, the input device 110 provides a GUI (Graphic User Interface) so that the user inputs necessary parameters rather than directly coding, which is an application with a neuromorphic architecture. development is possible

Therefore, in order to develop an application to which a neuromorphic architecture is applied, less time is required compared to the time required by a developer to directly code, thereby maximizing development efficiency.

In addition, it has the advantage of lowering the difficulty of development because it can be put into application development even with basic knowledge beyond the conventional development environment in which a skilled developer is required for application development.

The processor 120 is composed of an encoder component 121 and a decoder component 122, and the encoder component 121 may provide a static-dynamic data conversion function.

The encoder component 121 may be configured to encode two-dimensional static data into dynamic data including time intervals based on parameters input from the input device 110 (FIG. 2).

In order to apply the static data used in the existing artificial neural network to the spiking neural network, it is necessary to convert it into dynamic data containing time series information. This operation is called encoding, and encoding is performed through an encoder.

The present invention provides three conversion methods, and data conversion can be controlled by setting parameter values. The time interval, which is a parameter value, adjusts the firing time of the spike signal, and as the time interval increases, the delay between the nth spike and the n+1th spike increases.

At this time, the two-dimensional image data has the number of pixels of N*M according to the size of the image, and each pixel is composed of a value between 0 and 255. The dynamic data converted through the encoding process is converted from input 2D data to continuous time-series data in the form of 3D tensor to which time intervals are applied. Time series data consists of a value of 1 or 0 for each time interval, and represents the firing probability of a neuron at each time point.

The encoder component 121 may use one or more selected from a simple encoding method, a Poisson encoding method, and a population encoding method.

As shown in FIG. 3, in the simple encoding method, a spike value corresponding to a pixel exceeding a threshold value is obtained by passing an input image through a filter value of the same size randomly generated. Creates a filter with the same size as the image and randomly generates a value corresponding to each pixel in the image on a one-to-one basis.

The range of filter values can be adjusted, and a spike train having stochastic characteristics can be obtained as the threshold value of the filter increases.

At this time, the spike refers to an electrical wave generated when biological neurons exchange information. The SNN model starts with three assumptions based on biological facts. First, artificial nerve cells receive many inputs and output one spike signal. Second, the probability of generating a spike increases as the input becomes larger and decreases as the input becomes weaker. Third, it determines whether spikes are generated with only one reference point. Spike's mathematical model is the Dirac-Delta function. Intermittent spikes that occur over time are expressed as the sum of Dirac-delta functions corresponding to each spike, and this sum is expressed as a spike train.

As shown in FIG. 4, the Poisson encoding method generates spike data to which a Poisson distribution is applied based on the input intensity. It requires a firing rate in non-negative Hz units, and the interval between spikes cannot be zero. After receiving the input data size and pixel value, the firing speed is calculated in seconds using the intensity function.

Then, a Poisson distribution is generated through the calculated result, and the interval between spikes is sampled. To avoid the zero interval, the interval is increased by 1, and the spike time is calculated by accumulating and summing the spikes for each interval over time.

As shown in FIG. 5, in the population encoding method, when the number of neurons is input, the responses of neurons having different response characteristics to stimuli are linearly combined among different neurons in a cluster vector method. way to express it.

The average value of a plurality of responses includes the form of a Gaussian tuning curve that changes linearly according to the stimulus intensity, and responds most strongly to a stimulus similar to the average.

Through the encoder component 121, a function of converting static data such as an image of size N*M into dynamic data including time-series information may be provided, and at least one of three methods may be provided according to the spiking neural network model. can be provided to maximize compatibility.

In the present invention, static data may be converted into spike data having different shapes over time using the encoder component 121 .

As shown in FIG. 6, the decoder component 122 is a function of restoring dynamic data generated by the encoder component 121 into static data for identification and accuracy determination.

As a result of the operation of the spiking neural network model, data converted from inputs such as images and voices rather than class classification may be returned, and time-series data in the same spike form as the inputs may be generated.

The time-series data generated by the Spiking Neural Network model can be used as an input to other Spiking Neural Network models, and can be used as an input to IoT or artificial intelligence components.

In addition, if it is not the same spiking neural network, it must be restored in the same format as the input data before being converted to original time series data.

According to the present invention, as a result of performing decoding using the parameter values used in the encoding process, static data having a high similarity to the original data can be obtained compared to a result without using the parameter values.

In addition, according to the neural coding-based conversion method, the input spike row data is classified using the same different parameters used in the encoder.

As a result of the operation of the decoder component using the converted spike column data using the MNIST dataset as an input, it can be converted into static data similar to the original data used in the encoder component.

The decoder component 122 may provide a function of restoring the converted dynamic data to static data using the same parameter values used in the encoder component 121 for identification and accuracy determination.

The present invention can support both static and dynamic data-based application development supporting a spiking neural network model by using an encoder component and a decoder component that provide a static-dynamic data conversion function.

110: input device
120: processor
121: encoder component 122: decoder component
130: storage

Claims (6)

an input device 110 that receives component information;
a processor 120 processing component information input from the input device; and
A memory device 130 for storing component information processed by the processor 120,
The processor 120 is a spiking neural network input data processing system of a GUI-based integrated development environment, characterized in that composed of an encoder component 121 and a decoder component 122.
According to claim 1,
The encoder component 121 is a spiking neural network input data processing system of a GUI-based integrated development environment, characterized in that it provides a static-dynamic data conversion function.
According to claim 2,
The input device 110 is
A spiking neural network input data processing system of a GUI-based integrated development environment, characterized in that it provides a graphical user interface (GUI).
According to claim 3,
The encoder component 121 is
A spiking neural network input data processing system of a GUI-based integrated development environment, characterized in that for encoding two-dimensional static data into dynamic data including a time interval based on the parameters input from the input device 110.
According to claim 4,
The encoder component 121 is
A spiking neural network input data processing system of a GUI-based integrated development environment, characterized by using one or more encoding methods selected from simple encoding method, Poisson encoding method, and population encoding method.
According to claim 5,
The decoder component 122 is
A spiking neural network input data processing system of a GUI-based integrated development environment, characterized in that the encoded dynamic data is reconverted into static data.

Images