KR102553366B1 - Spiking Neural Network-Based Component Generating System for Intelligent Application IoT - Google Patents

Abstract

The present invention relates to a spiking neural network-based component generation system for intelligent application IoT, and more particularly, by providing an IoT developer with a function of generating an intelligent component including a Spiking Neural Network (SNN) model compatible with an IoT platform, It is about a component generation system that can be linked with neuromorphic hardware as well as applying the SNN model of IoT applications.

Description

Spiking Neural Network-Based Component Generating System for Intelligent Application IoT}

The present invention relates to a spiking neural network-based component generation system, and more particularly, to IoT developers by providing a function of generating an intelligent component including a Spiking Neural Network (SNN) model compatible with an IoT platform, thereby providing an SNN of existing IoT applications. It is about a component generation system that can be linked with neuromorphic hardware as well as model application.

IoT (Internet of Things) is a core concept of the 4th industrial revolution, meaning that all things are connected to the Internet. In other words, information is integrated and managed by connecting various smart devices used in smart phones, home appliances, healthcare, self-driving cars, smart homes, smart factories, and smart farms through the Internet.

IoT has developed from the limited function of the connected IoT that conveys the state of things, and is developing in the direction of communicating with other IoT devices, recognizing the surrounding situation, and even autonomously responding.

In particular, the hardware-based approach to improve the performance of IoT edge services has a strong prospect that data processing using neuromorphic hardware will be possible. Therefore, it has become essential to support a framework combined with various hardware, including neuromorphic hardware.

At this time, the neuromorphic hardware uses a neural network structure, which refers to a structure adopted by computers to solve problems in a similar way to how humans process problems through the brain.

Just as neurons, the basic structural organization of the human brain, are connected to each other to process work, when neurons as a mathematical model are interconnected to form a network, it is called a neural network. It is also called an artificial neural network to distinguish it from a biological neural network.

A neural network has excellent parallelism because each neuron acts as a processor that operates independently, and since information is distributed over many connection lines, even if a problem occurs in a few neurons, it does not significantly affect the entire system, so it has fault tolerance. and has the ability to learn about a given environment.

together. SNN (Spiking Neural Network), the latest model of neural networks, is a next-generation artificial neural network model that introduces biological concepts to existing neural networks. SNN is a 3rd-generation artificial neural network model that mimics the spike delivery and processing of neurons, which are brain nerve cells, and is a model that solves the problems of existing deep learning models.

However, conventional IoT frameworks do not support machine learning using such neural networks. Moreover, it is difficult to utilize it because it is not compatible with neuromorphic hardware for supporting SNN models. In addition, there was a problem that it was difficult for IoT developers to understand the inside of the SNN framework due to its complex structure and parameters, making it difficult to apply it to IoT. Due to these problems, there is no application of SNN framework in existing IoT applications.

In order to solve these problems, the need for development tools that provide compatibility between IoT applications and artificial neural network model frameworks has been gradually emerging.

Korean Patent Publication No. 10-2017-0094355 (published on August 17, 2017)

The present invention, conceived to solve this problem, relates to a spiking neural network-based component generation system for intelligent application IoT, and more specifically, provides IoT developers with a Spiking Neural Network (SNN) model compatible with the IoT platform. It is about a component creation system that provides a function of generating intelligent components and can be linked with neuromorphic hardware as well as applying the SNN model of existing IoT applications.

Spiking neural network-based component generation system for intelligent application IoT according to an embodiment of the present invention for achieving the above object,

a parameter management unit 100 that acquires essential parameters of neurons; and

It may be configured with a model generation unit 200 that converts the model generated by the parameter management unit 100 into code of an execution environment to generate a component.

At this time, the parameter management unit 100 includes a neuron model parameter acquisition unit 110 for acquiring neuron model definition information, a training parameter acquisition unit 120 for acquiring neural network learning information, and a layer environment acquisition unit for acquiring neural network layer information ( 130), but

The interface of the parameter management unit 100 may be composed of a GUI.

In addition, the component generated by the model generation unit 200 is a basic information file 211 including the model name, model version, and information of the model maker of the generated component, and an executable file of the generated component. A component file 210 composed of a function execution file 212 defining

When machine learning is applied as an implementation file of the generated component, it may be composed of an application file 220 that performs a neural network operation.

In addition, in the component file 210, the basic information file 211 and the function execution file 212 may be compatible with the IoT framework.

In addition, the application file 220 further includes a platform compatibility providing means 221 providing compatibility with an external machine learning platform and a hardware compatibility providing means 222 providing compatibility with neuromorphic hardware. It can be.

At this time, the model generating unit 200 may generate an IoT support and application component, an Artificial Neural Network (ANN) support component, and a Spiking Neural Network (SNN) support component.

The spiking neural network-based component generation system for the intelligent application IoT of the present invention provides IoT developers with biological knowledge of the SNN model in creating applications using SNN, a next-generation artificial neural network model that requires biological knowledge of a complex structure in a device. It improves the development environment for IoT developers by receiving necessary values through a simple GUI and automatically generating intelligent components for IoT applications without knowledge of structures and complex processing processes, and provides more AI services to IoT applications. Easy to apply.

In addition, compatibility with various platforms such as PyTorch, Caffe2, TensorFlow, etc. is provided through platform compatibility means to provide a more open development environment to IoT developers, thereby increasing the efficiency of the IoT application program.

In addition, it is possible to provide an environment in which IoT applications can perform more efficient machine learning by providing compatibility with neuromorphic hardware such as NengoDL and NAAL by providing Native Code through a hardware compatibility providing means.

1 is a system configuration diagram of the present invention
2 is a configuration diagram of a parameter management unit of the present invention
3 is an embodiment of a parameter management unit of the present invention
4 is a block diagram of the model generation unit 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.

1 is a system configuration diagram of the present invention, FIG. 2 is a configuration diagram of a parameter management unit of the present invention, FIG. 3 is an embodiment of a parameter management unit of the present invention, and FIG. 4 is a configuration diagram of a model generation unit of the present invention.

As shown in FIG. 1, the spiking neural network-based component generation system for intelligent application IoT of the present invention,

It can be composed of a parameter management unit 100 that acquires essential parameters of neurons for IoT developers, and a model generation unit 200 that converts the models generated by the parameter management unit into codes of an execution environment to create components.

At this time, the parameter management unit 100 includes a neuron model parameter acquisition means 110 for acquiring neuron model definition information, a training parameter acquisition means 120 for acquiring neural network learning information, and a layer environment acquisition for acquiring neural network layer information. It may consist of means (130).

That is, conventionally, in order for an IoT developer to add a function using a neural network platform such as machine learning to an IoT application, knowledge of the deep biological structure, operation, and processing of a neural network model is additionally acquired and the application of the corresponding neural network model is implemented. components for the desired platform had to be produced by hand.

This inevitably had an adverse effect of increasing the entire development time of the IoT application by making IoT developers spend a long time learning and developing.

However, the spiking neural network-based component generation system for the intelligent application IoT of the present invention provides the latest neural network model, SNN (Spiking Neural Network) An IoT application to which the SNN model is applied can be developed without additional work by generating components necessary for application of the model in the model generator 200.

In particular, the neuron model parameter acquisition means 110, the training parameter acquisition means 120, and the layer environment acquisition means 130 request essential information necessary for defining a neuron model from an IoT developer, and a neuron model reflecting the corresponding information. By receiving detailed information for definition and environment setting, IoT developers can create neuron models optimized for the IoT application they are developing.

Therefore, IoT developers can more easily apply SNN models to IoT applications, enabling various uses, additionally acquiring knowledge about the deep biological structure, operation, and processing process of neural network models, and development for compatibility. It can shorten the development period of the entire IoT application program.

In addition, using these characteristics, it implements the Core Neural Network library, measures its performance, and applies it to various open IoT frameworks to support edge computing application development using neuromorphic architecture hardware that improves energy consumption and latency efficiency. possible.

At this time, the interface of the parameter management unit 100 may be composed of a GUI. Therefore, it is possible to develop an application program more quickly and accurately by minimizing the human error factor of the developer that may occur when developing using an interface composed of text only.

In addition, the interface of the parameter management unit 100 may provide a wireframe type interface. Through this, IoT developers can more easily recognize the setting of the neural network, and can more easily recognize the overall characteristics of the neural network model, increasing work efficiency.

In addition, as shown in FIGS. 1 to 4, the components generated by the model generation unit 200 include a basic information file 211 including the model name, model version, information of the model maker, and the like of the generated component. The component file 210 composed of the function execution file 212 defining the operation of the model as an executable file of the generated component, and the application file that performs the neural network operation when machine learning is applied as the implementation file of the generated component ( 220).

That is, the file generated by the model generator 200 can be divided into the component file 210 and the application file 220 .

At this time, the component file 210 is the basic information file 211 containing basic production information of the model, such as model name, model version, and information of the model maker, and the execution file of the component, which defines the operation of the model. It is composed of a function execution file 212 that does. In addition, in the component file 210, the basic information file 211 and the function execution file 212 may be compatible with the IoT framework.

Therefore, on a platform such as Node RED, machine learning can be performed by checking the information of the neural network model through the basic information file 211 and checking the operation through the function execution file 212. Here, Node RED is a flow-based programming tool that composes flows containing nodes. It is combined with the Node.js engine using Python and Node.js to provide a web browser-based editor, and developers can easily drag and drop nodes from the palette. It is a platform for building applications.

In addition, the model generation unit 200 may further include a sandbox capable of testing the operation of the generated neural network model.

That is, it is possible to generate a neural network model through the model generating unit 200, and test whether the generated neural network model implements the function intended by the IoT developer and is connected to an application program and properly operated. Therefore, it is possible to prevent an increase in the development period due to a problem being discovered in the middle of development by conducting a test in advance to the IoT developer to check for problems in advance.

In addition, the application file 220 further includes a platform compatibility providing means 221 providing compatibility with an external machine learning platform and a hardware compatibility providing means 222 providing compatibility with neuromorphic hardware. It can be.

The platform compatibility providing means 221 can make the components of the present invention compatible with different neural network framework environments such as PyTorch, Caffe2, and Tensorflow.

In general, in order to use a specific neural network framework environment, it was necessary to work according to the corresponding neural network framework, but through the platform compatibility providing means 221, the spiking neural network-based component generation system for the intelligent application IoT of the present invention is an IoT developer It can provide compatibility according to the neural network framework set in advance. Through this, IoT developers can secure compatibility for each platform without separate work, maximizing work efficiency.

In addition, native code is provided through the hardware compatibility providing means 222 to provide compatibility with neuromorphic hardware such as NengoDL and NAAL, thereby providing an environment in which IoT applications can perform more efficient machine learning.

In addition, the model generation unit 200 may generate IoT support and application components, ANN (Artificial Neural Network) support components, and SNN support components. Through the characteristics of the model generation unit 200 as described above, it is possible to generate components corresponding to various formats, thereby increasing the efficiency of application program development by IoT developers.

As described above, the present invention has been described with specific details such as specific components and limited embodiment drawings, but these are provided to help a more general understanding of the present invention, and the present invention is not limited to the above embodiments, Those skilled in the art in the field to which the present invention pertains can make various modifications and variations from these descriptions.

Therefore, the spirit of the present invention should not be limited to the described embodiments, and not only the claims to be described later, but also all modifications equivalent or equivalent to the scope of the claims belong to the scope of the present invention.

100: parameter management unit
110: means for acquiring neuron model parameters
120: training parameter acquisition means
130: layer environment acquisition means
200: model generation unit
210: component file
211: basic information file 212: function execution file
220: application file
221: means for providing platform compatibility 222: means for providing hardware compatibility

Claims (6)

a parameter management unit 100 that acquires essential parameters of neurons; and
In the spiking neural network-based component creation system for intelligent application IoT including;
The parameter management unit 100
Neuron model parameter acquisition means 110 for acquiring neuron model definition information;
Training parameter acquisition means 120 for acquiring neural network learning information;
Consists of a layer environment acquisition means 130 for acquiring neural network layer information,
The components created by the model generating unit 200 are
A basic information file 211 including the model name of the generated component, the model version, and information of the model maker, and
A component file 210 composed of a function execution file 212 defining the operation of the model as an executable file of the generated component;
It is an implementation file of the generated component and consists of an application file 220 that performs neural network work when machine learning is applied.
The application file 220 is
A platform compatibility providing means 221 for providing compatibility with an external machine learning platform;
Includes hardware compatibility providing means 222 for providing compatibility with neuromorphic hardware;
The model generator 200
Create IoT support and application components, ANN (Artifical Neural Network) support components and SNN support components,
Develop an IoT application to which a neural network model is applied by inputting essential values for forming a neural network model into the parameter management unit 100 and generating components necessary for applying the neural network model in the model generating unit 200,
Through the platform compatibility providing means, compatibility with various platforms is provided to increase the efficiency of the IoT application program,
A spiking neural network-based component generation system for intelligent application IoT, characterized in that the IoT application program performs efficient machine learning by providing compatibility with neuromorphic hardware through the hardware compatibility providing means.
delete delete According to claim 1,
The component file 210 is
Spiking neural network-based component creation system for intelligent application IoT, characterized in that the basic information file 211 and the function execution file 212 are compatible with the IoT framework.
delete delete

Images