Conceptio › Archive › Google Patents

Google Patentsopen access

Artificial-intelligence decision-making core system with neural network — Ahp-Tech Inc. (US11526726B2)

Ahp-Tech Inc. · Google Patents

Google Patents · Patents · License: Open Access

Open Source ↗

chao

patent, google patents, intellectual property, US11526726B2, Ahp-Tech Inc., Chao-Huang CHEN, en, 2022

ABSTRACT

Abstract

An artificial-intelligence decision-making core system with neural network implements asymmetric hidden layers which is constructed by neural network with a dynamic neuron adjusting mechanism via making use of a device with computing unit and storage media, coupled to an independent feedback sub-system which operates a residual-compensation mechanism, thereby the core system can receive various trained data and performs a non-linear analysis process according to the output data derived from the asymmetric hidden layers of an unsupervised neural network, so as to derive individual and applicable decision-making data.

Description

CROSS-REFERENCE TO RELATED APPLICATION

This non-provisional application claims priority under 35 U.S.C. Â§ 119(a) on Patent Application No(s). 109122440 filed in Taiwan, R.O.C. on Jul. 2, 2020, the entire contents of which are hereby incorporated by reference.

BACKGROUND OF THE INVENTION

1. Field of the Invention

The present invention relates to a decision-making core system, and in particular, to an artificial-intelligence decision-making core system with neural network to dynamically adjust neuron nodes, make use of non-linear analysis processes, and couple residual-compensation mechanism, thereby the core system can support a variety of trained data with various properties to derive decision-making data with differentiation and applicability.

2. Description of the Related Art

At present, most of conventional artificial-intelligence decision-making systems are applied with a single type of neural-network topology, and the data structures adopted by the neurons arranged and configured in neural networks are usually traditional linear data structures.

However, the above-mentioned conventional decision-making system with the single neural network which adopts linear data structure for data training, usually derives non-distinguished training results. As expected, decision-making systems with similar neural-network topologies normally work in similar decision-making styles as well. It makes the decision-making system not readily to be applied to different scenarios, and cannot provide decision-making results with different variable dimensions or styles for advanced commercial applications. Moreover, it will degrade the effect of decision making due to worse reference and applicability of the decision-making results.

Even though there are already a few types of neural networks can provide decision-making technology with differentiation on the output results (such as DCIGN, LSM, SOMNN, etc.), once the topology of neural network is trained to fit data with certain features (such as financial data, human factors data, chess data, etc.), it is difficult to change the neural networks to apply to data with different features. Moreover, its data of decision-making results is also easy to be learned or analyzed by an external neural network to crack the topology structure of its neural network.

Therefore, it is desirable to have improvements in the approach to artificial-intelligence decision-making systems.

BRIEF SUMMARY OF THE INVENTION

In view of the drawbacks of the prior art, it is an object of the present invention to provide an artificial-intelligence decision-making core system with neural network, wherein the system can dynamically adjust neuron nodes, make use of non-linear analysis processes, and couple a residual-compensation mechanism, thereby the system can support a variety of trained data with various properties to derive individual and applicable decision-making data. The related technologies can be implemented as a device or an electronic system with reasonable cost, and can quickly adjust the topology of the neural network according to the specific requirements for decision-making data, while effectively avoids the defect of the prior art systems where their output data is easy to be learned or analyzed by external neural networks.

To achieve the above objects and more, the present invention provides a device or system to implement the decision-making core system. The present invention comprises an artificial-intelligence decision-making core system with neural network comprising: an unsupervised neural-network interface module, a neuro-computing sub-system, and a residual-backpropagation sub-system. The unsupervised neural-network interface module comprises at least a computing-and-storage device, wherein this device can be implemented as an input device such as a computing device to receive raw data, and an interface of this input device for receiving the raw data can be applied to a variety of trained data from external non-specific models (such as financial models, human behavior models, war chess models, etc.). The neuro-computing sub-system comprises at least a computing device integrated with or coupled to at least one storage device, wherein the sub-system is connected to the above-mentioned input device to perform pre-processing operations on the received raw data to obtain pre-processed data, wherein the pre-processing operations are performed in an asymmetric-hidden-layers input module within an embedded system, and then the pre-processed data is subjected to a neuron computing process that can dynamically update the neuron nodes, and then the neuro-computing sub-system is capable of deriving corresponding non-linear analyzed data and decision-making data, via a non-linear PCA (principle component analysis) computing process with assistance of Chaos models. The residual-backpropagation sub-system that can operate independently, coupled to the neuro-computing sub-system, is used for performing a residual compensation computing process based on the non-linear analyzed data to generate system tuning information and residual-feedback data, or training-breaker information, to facilitate the neuro-computing sub-system to derive individual and applicable decision-making data, wherein the system tuning information is coupled to the neuro-computing sub-system for modifying or adjusting the neural network, and the training-breaker information is coupled to the neuro-computing sub-system for performing a breaking mechanism for the training processes (hereinafter training-break mechanism) in the neural network, and the residual-feedback data is for facilitating the neuro-computing sub-system to derive or enhance decision-making data.

In an embodiment, the above-mentioned neuro-computing sub-system comprises: an asymmetric-hidden-layers input module, a tree-structured neuron module, a layered weight parameter module, a non-linear PCA module, and a tree-searching module. The asymmetric-hidden-layers input module is configured to perform a data pre-processing process on the raw data received from the unsupervised neural-network interface module, in asymmetric hidden layers formed with an unsupervised neural network, to derive pre-processed data, wherein the mentioned asymmetric hidden layer means that the bidirectional weights between neurons are not equal in such hidden layer. The tree-structured neuron module, connected to the asymmetric-hidden-layers input module, including a plurality of neuron-node data, is used for managing neurons of a neural network, wherein the module can perform a tree-structured data processing program based on the pre-processed data to generate tree-structured data corresponding to each neuron node, whereby the neuron nodes can be modified according to the system tuning information from the residual-backpropagation sub-system The layered weight parameter module, connected to the tree-structured neuron module, is configured to perform a layered weight parameter computing process according to the tree-structured data to obtain layered-weight parameter data for managing the specific weight of each hidden layer in the neural network, wherein it makes considering the factors of the layered weight parameter be necessary in addition to the weight parameters of individual neuron, for tuning the weight parameter of the neural network. The non-linear PCA module, connected to the layered weight parameter module, is configured to perform a non-linear PCA computing process based on the layered-weight parameter data to derive non-linear analyzed data, wherein the non-linear PCA module is configured to derive applicable decision-making data based on the non-linear analyzed data or work with the residual-feedback data from the residual-backpropagation sub-system for improving the decision-making data. The tree-searching module, coupled to the tree-structured neuron module and the residual-backpropagation sub-system, is utilized for breaking the neural-network training process at once to maintain the differentiation and the applicability of the decision-making data. Combining these elements in an implementation of software, hardware, or a combination of software and hardware as a neuro-computing sub-system can dynamically adjust neuron nodes to support various types of trained data, so that the unsupervised neural network with asymmetric hidden layers can export layered-weight parameter data covering multi-dimensional variables, and can also import applicable residual-compensation mechanism dynamically, whereby a set of data with non-linear multiple variables can be further eigenized through a non-linear PCA computing process based on the layered-weight parameter data, to make it practical to support various data types within a decision-making neural network. It is an effect that cannot be achieved by the prior art.

In order not to cause any misunderstanding or misleading in the following descriptions for the present invention, please notice that in the above embodiment, the non-linear PCA computing process is not based on the data derived from the output layer of a neural network. Instead, the non-linear PCA computing process is performed according to the data related to the layered weight of each hidden layer in a neural network. A legacy artificial-intelligence application usually treats the hidden layers as a black box within a neural network, but the present invention analyzes and eigenizes how the neural network makes decisions through the hidden layers. This technology idea is much progressive and creative from the conventional artificial-intelligence applications.

In an embodiment, in order to enhance the decision-making quality of the core system, the above-mentioned neuro-computing sub-system further comprises the following modules: a hidden-layered routing module, a dense data processing module, a sparse tensor processing module, a Chaos model analysis module, and a neuron dimension switching module. The hidden-layered routing module receives the raw data from the unsupervised neural-network interface module of the decision-making core system, and plans a route (e.g., a feasible route or so on) for the hidden layers according to the attributes of the raw data, so that the neuro-computing sub-system can support variable propagation paths for the hidden layers of the neural network dynamically. The dense data processing module analyzes the raw data from the unsupervised neural-network interface module of the decision-making core system and performs a dense data processing program for dropping invalid data if there is a high-density attribute with the raw data, after that normalizes the duplicate data, and then import the output data which is already completed by the dense data processing program, into the asymmetric-hidden-layers input module of the neuro-computing sub-system. The sparse tensor processing module, coupled to the tree-structured neuron module of the sub-system, analyzes the neuron nodes derived by the tree-structured neuron module to verify whether the neuron nodes fit the conditions of sparse tensor, wherein when the neurons fit the condition, the sparse tensor processing module performs a sparse-matrix operation process to facilitate computing performance and save system memory usage for the neural network. When the non-linear PCA module cannot make effective corrections according to the residual-feedback data received from the residual-backpropagation sub-system, the Chaos model analysis module, coupled to the non-linear PCA module of the sub-system, performs a Chaos model analysis process to facilitate the derivation of the decision-making data. The neuron dimension switching module, coupled to the tree-structured neuron module, performs a neuron-dimension switching process based on system tuning information from the residual-backpropagation sub-system, to facilitate the switching of the decision-making style, so that the neuron dimension switching module is capable of enhancing the differentiation and applicability of the decision-making data. Combining these elements in an implementation of software, hardware, or a combination of software and hardware into a neuro-computing sub-system can facilitate the derivation of the decision-making data and enhance the performance of derivation, and can make real-time decision-making in practice. This integration is much progressive from the prior art.

In an embodiment, the above-mentioned residual-backpropagation sub-system comprises: a progressive backpropagation input module, a system tuning module, and a progressive backpropagation output module. The progressive backpropagation input module is for receiving non-linear analyzed data from the above-mentioned non-linear PCA module of the neuro-computing sub-system. The system tuning module is configured to perform a residual compensation process for the non-linear analyzed data, to derive a feedback-tuning information, and to trigger a training-break mechanism of neural network according to the feedback-tuning information or deliver the feedback-tuning information. The progressive backpropagation output module, based on the updated state of the system tuning module and the derived feedback-tuning information, is configured to derive a residual-feedback data and send the residual-feedback data to the non-linear PCA module of the neuro-computing sub-system for facilitating the neuro-computing sub-system to maintain the differentiation and applicability of the decision-making data. Combining these elements in an implementation of software, hardware, or a combination of software and hardware as a residual-backpropagation sub-system not only improves the accuracy of the neural network training model, but also can effectively manage the distribution interval of decision-making data, and can also timely avoid the training model of the neural network being hacked by unsuitable external data. This integration is much progressive from the conventional technologies.

In an embodiment, in order to enhance the processing ability for non-linear analyzed data, so as to apply various non-linear analyzed data with different properties, wherein the residual-backpropagation sub-system further comprises: a dynamic-norm processing module, an inner-product processing module, a manifold operation module, a tensor switching module, a functional operation module, and an ARIMA-model processing module. The dynamic-norm processing module for non-linear analyzed data that can be vectorized or matrixed, can work with the progressive backpropagation input module of the residual-backpropagation sub-system to perform a pre-processing program to import the vectorized data or matrixed data to the system tuning module for operation of residual-compensation mechanism. In addition, if the non-linear analyzed data reveals that most of the layered weights are in a large scale, the progressive backpropagation input module of the residual-backpropagation sub-system can further coordinate the dynamic-norm processing module to perform a re-initialization process to implement a norm-penalty mechanism for the hidden layers to re-initialize the related weight parameters of the neural network for reducing possible overfitting. The inner-product processing module for vectorized data or matrixed data with image properties can work with the progressive backpropagation input module of the residual-backpropagation sub-system to perform another pre-processing program to self-define the relationship between nodes and vectors, so as to introduce standard orthogonal basis to the system tuning module to improve the operation of residual-compensation mechanism. The manifold operation module for non-Euclidean space data can work with the progressive backpropagation input module of the residual-backpropagation sub-system to perform another pre-processing program to switch the non-Euclidean space data into the form of Euclidean space, so as to import sampling data from various geometry spaces to the system tuning module to enhance the applicability of residual-compensation mechanism. The tensor switching module for time-variant data can work with the system tuning module of the residual-backpropagation sub-system to perform a self-correction process to switch the residual data already processed by residual-compensation mechanism through various dimensional spaces, so as to find the optimal residual-operation space that is most feasible for the time-variant data to enhance the applicability of residual-compensation mechanism. The functional operation module for a computing process of obtaining the optimal solution of loss function via making use of boosting algorithm can work with the system tuning module of the residual-backpropagation sub-system to perform a mathematical programming process with functional operations, to facilitate the derivation of the optimal solution of the loss function, so as to maintain the accuracy of the residual-compensation mechanism. The ARIMA-model processing module for non-stationary residual data can work with the system tuning module of the residual-backpropagation sub-system to perform a stationary process through integrated finite-difference operations to obtain a stationary residual data, so as to check whether the stationary residual data fits a white-noise sequence to confirm whether the original non-stationary residual data needs to be corrected. Combining these elements in an implementation of software, hardware, or a combination of software and hardware into a residual-backpropagation sub-system not only facilitates residual operation for data with various properties, b

CROSS-REFERENCE TO RELATED APPLICATION

BACKGROUND OF THE INVENTION

1. Field of the Invention

2. Description of the Related Art

Therefore, it is desirable to have improvements in the approach to artificial-intelligence decision-making systems.

BRIEF SUMMARY OF THE INVENTION

In an embodiment, the above-mentioned tree-structured neuron module of the neuro-computing sub-system comprises: a cross-neuron computing unit and a neuron-weight updating unit. The cross-neuron computing unit performs a neuron data updating process according to the residual-feedback data received by the neuro-computing sub-system from the progressive backpropagation output module of the residual-backpropagation sub-system, so as to generate corresponding neuron updating data. The neuron-weight updating unit performs another neuron data updating process according to the neuron updating data, so as to update the tree-structured data and the weight parameter data for the neurons of the neural network, and notify the asymmetric-hidden-layers input module to perform an updating process for the raw data in a pre-processing process. Combining these elements in an implementation of software, hardware, or a combination of software and hardware as a tree-structured neuron module not only provides the effect of real-time weight adjustment for the neural network, but also facilitates to dynamically adjust the neuron nodes of a neural network, so as to change its decision-making style in real time. This is an effect that cannot be achieved by conventional technology.

In an embodiment, in order to prevent being hacked or learned via external neural networks, the tree-structured neuron module of the neuro-computing sub-system further comprises: a neuron generator unit, a neuron topology switching unit, and a neuron grouping management unit. The neuron generator unit performs a neuron generating process according to the instructions from the tree-structured neuron module, so as to generate the required nodes for corresponding neurons. The neuron topology switching unit performs a neuron topology switching process according to the instructions from the tree-structured neuron module, working with the neuron generator unit, so as to dynamically change the topology structure of the neural network. The neuron grouping management unit performs a neural-network reorganizing process for neuron nodes with various attributes, according to the instructions from the tree-structured neuron module, working with the neuron topology switching unit, so as to dynamically combine multiple neural networks with different topology structures. Combining these elements in an implementation of software, hardware, or a combination of software and hardware into a tree-structured neuron module can achieve the effect of deploying various neural networks of different topology structures rapidly, so that the present invention can quickly switch between neural networks with different functions (for example: switching between supervised and unsupervised neural networks, switching between memorized and non-memorized neural networks, or switching between self-defined and open-sourced neural networks). This is an effect that cannot be achieved by conventional technology.

In an embodiment, the above-mentioned layered weight parameter module of neuro-computing sub-system comprises: an activation-function integration unit, a Laplace-transformation processing unit, and an activation-function updating unit. The activation-function integration unit is configured to work with various activation functions to perform a setting process according to a configured data of activation function managed by the layered weight parameter module and come out a setting result, so as to set an applicable activation function to a layered-weight parameter updating process for the neural network. The Laplace-transformation processing unit is configured to perform a Laplace transformation process according to the activation function configured by the layered weight parameter module and the weight-parameter data of the neurons to generate a Laplace operation result. The activation-function updating unit is configured to perform an activation-function updating process to derive a configured data of activation function corresponding to the layered-weight parameter data according to the Laplace operation result, wherein the layered weight parameter module can analyze and evaluate whether to coordinate the activation-function integration unit to reconfigure the corresponding activation function. Combining these elements in an implementation of software, hardware, or a combination of software and hardware into a layered weight parameter module with a storage media which is applicable for storing data of neuro-weight parameter and configured data of activation function, not only provides various activation functions for a neural network so that the neuro-computing sub-system can switch between neural networks with different functions, but also can quickly confirm whether the configured activation function is feasible for a time-variant data set. This is an effect that cannot be achieved by conventional technology.

In an embodiment, in order to enhance the generalization ability of the neural network, the above-mentioned layered weight parameter module of the neuro-computing sub-system further comprises: a normalization processing unit, a harmonic-function processing unit, a time-variant function processing unit; a multi-dimensional moment processing unit, a Fourier-transformation processing unit, and a wave-function processing unit. The normalization processing unit is driven for optimizing the performance of hidden layers according to the setting result from the activation-function integration unit, to perform a normalization processing program for an unbounded activation function. In addition, the harmonic-function processing unit is driven for the raw data that has phasing properties or can be expressed on a complex plane according to the setting result from the activation-function integration unit, to perform a harmonic-function mapping process, so that the configured activation function can support data which is in the form of complex number. Moreover, the time-variant function processing unit is driven to import time variables to the configured activation function if the layered weight parameter module determines that the hidden layer is unable to avoid gradient vanishing problem or gradient exploding problem according to the configured data of activation function, so as to facilitate the result of weight adjustment tending to converge. The multi-dimensional moment processing unit is driven to introduce an affine transformation process for a moment generating function to the configured activation function according to the configured data of activation function if the layered weight parameter module determines that the raw data fits the property of multi-variant continuous random variables, so as to facilitate the result of weight adjustment tending to converge. The Fourier-transformation processing unit configured to be driven to perform a Fourier-transformation process on the data propagated by the hidden layers via making use of the Dirac delta function to obtain a transformed data if the layered weight parameter module determines that the raw data has the distribution property of the Dirac delta function according to the configured data of activation function, so that the Fourier-transformation processing unit uses a ramp function as an activation function to import the transformed data into the activation function, for supporting the raw data with particle properties or the raw data in zero dimension. The wave-function processing unit is driven to perform a wave-function processing program to use an applicable wave function as the configured activation function according to the configured data of activation function if the layered weight parameter module determines that the raw data has the distribution property of a quantum many-body system, so as to facilitate the result of weight adjustment for quantum many-body system simulation tending to converge. Combining these elements in an implementation of software, hardware, or a combination of software and hardware into a layered weight parameter module not only enhances processing performance of the hidden layers, but also can facilitate whether the configured activation function is feasible for a time-variant data set. This is an effect that cannot be achieved by conventional technology.

In an embodiment, the above-mentioned tree-searching module of the neuro-computing sub-system comprises: a duplicate-neurons preventing unit and an upper confidence bounds applied to trees (UCT) searching unit. The duplicate-neurons preventing unit is configured to determine whether there are duplicate neuron nodes according to the tree-structured data derived by the tree-structured neuron module of neuro-computing sub-system, wherein the duplicate-neurons preventing unit derives duplicated-neuron information if there are duplicate neurons, thereby the tree-searching module can drive the state register unit of the residual-backpropagation sub-system to derive a corresponding training-breaker information, so that the neuro-computing sub-system can coordinate the tree-structured neuron module of the neuro-computing sub-system according to the training-breaker information, to perform a training-break mechanism. The UCT searching unit which performs an upper confidence bound (UCB) computing process based on the system tuning information from the residual-backpropagation sub-system and the raw data, tree-structured data, weight-parameter data, and non-linear analyzed data of the neuro-computing sub-system, to generate corresponding upper-confidence-bound information, thereby the tree-searching module determines whether the current neural network does not fit the UCB information and coordinate the state register unit to derive a training-breaker information if it does not, so that the neuro-computing sub-system can coordinate the tree-structured neuron module of the sub-system according to the training-breaker information, to perform a training-break mechanism. Combining these elements in an implementation of software, hardware, or a combination of software and hardware into a tree-searching module not only prevents the invalid training program from consuming system resources, but also further manages the distribution interval of decision-making data, and avoids wasting too much system resources in unnecessary decision-making data analysis processes.

In an embodiment, in order to enhance the tree-searching performance, the above-mentioned tree-searching module of the neuro-computing sub-system further comprises: a support vector machine (SVM) processing unit, a genetic algorithm (GA) searching unit, a Monte Carlo tree searching unit, a searching-schedule accelerating unit, and a hash-table searching unit. The SVM processing unit is driven to perform a hyperplane optimization process if the tree-searching module determines the current upper-confidence-bound is out of an applicable range according to the upper-confidence-bound information from the UCT searching unit, so as to adjust the decision-making boundary and optimize the upper-confidence-bound for avoiding too many invalid searching processes. In addition, the GA searching unit is driven to perform a genetic-algorithm computing process for a neural network composed of a variety of topological structures, so as to group neuron nodes to determine a target group, thereafter to perform a tree-structured data searching process on the target group, so that the tree-searching module can effectively simplify the searching scope. Moreover, the Monte Carlo tree searching unit is driven to perform a Monte Carlo tree search (MCTS) process for a neural network based on a tree-structured topology with a large depth, so that the tree-searching module can enhance the tree-searching performance through heuristic-searching processes, for better paths. The searching-schedule accelerating unit is driven to perform a B+tree searching process in a file system (e.g., NTFS is more suitable) embedded with B+tree file structure, for a neural network based on a tree-structured topology with a large width, so that the tree-searching module can perform the searching schedule with B+tree searching algorithm implementing range queries and ordered iteration based on the bottom-most index blocks, which can effectively save the search time. The hash-table searching unit is driven to store indexes in a hash table and perform an index-searching process with hash-searching algorithm, for a neural network composed of a tree-structured topology with a large number of indexes, so that the tree-searching module can effectively save search time. Combining these elements in an implementation of software, hardware, or a combination of software and hardware into a tree-searching module not only effectively solves the problem of system searching performance which is the key issue for neuron nodes adopting tree-like data structure, but also avoids large amount of system memory occupation. In some fields, especially for those data related to a local curvature which can be expressed as a Hessian matrix mapping to a function of many variables, such combination as a tree-searching module can do it well in efficiency via MCTS, UCT, and GA related technologies within the present invention. These effects make the neurons with tree-structured data be feasible to the technology field of neural network in industrial utilization.

In another embodiment, the above-mentioned system tuning module of the residual-backpropagation sub-system comprises: a residual analysis unit, a self-tuning processing unit, a state register unit, and a training breaker unit. T residual analysis unit, with more than one residual computing program, is configured to perform an applicable residual computing program according to the non-linear analyzed data imported by the progressive backpropagation input module to generate a residual analysis data. The self-tuning processing unit is configured to perform a self-tuning processing program based on the non-linear analyzed data and the residual analysis data, with an adaptive residual analysis for a backpropagation of residual compensation, to derive a system tuning information, wherein the system tuning information is for fine-tuning the neural network or for triggering a training-break mechanism. For the state register unit, the system tuning module can store more than one break-configuration switching data in the state register unit according to the system-state data. The training breaker unit is configured to determine whether a training-break condition is met according to the system tuning information from the self-tuning processing unit, wherein the training breaker unit works with the state register unit if the training-break condition is met, to generate a training-breaker information according to a selected break-configuration switching data, so that the residual-backpropagation sub-system can further coordinate the neuro-computing sub-system to perform a training-break mechanism according to the training-breaker information. Combining these elements in an implementation of software, hardware, or a combination of software and hardware as a system tuning module not only facilitates the applicability of decision-making data via making use of proper system-tuning mechanism, but also effectively ensures the individuality of decision-making data through introducing the training-break mechanism, so as to avoid the issue that the legacy unsupervised neural network is easy to over-eigenize non-important features of data during a training process so that will tend to derive ordinary decision-making data.

In another embodiment, in order to enhance the management for training states and system tuning information, the above-mentioned system tuning module of the residual-backpropagation sub-system further comprises: a state recovery unit, a state switching unit, a data encoding unit, a state deletion unit, a visualized monitoring unit, a data compressing unit, and a history tracing unit. The state recovery unit is for a system that has already performed a training-break mechanism, wherein if the factor that caused the training-break has been eliminated, the state recovery unit will be driven to facilitate the system to restore a normal state which is before the training-break mechanism is triggered, according to the break-configuration switching data stored in the state register unit. The state switching unit is driven to perform a state switching process according to the break-configuration switching data stored in the state register unit if the current decision-making result is more degraded than the previous decision-making result or there are multiple types of raw data that must be switched, so as to facilitate the system to switch to a previous state which is more applicable. In addition, the data encoding unit is for enhancing the confidentiality of system tuning information, residual-feedback data, or training-breaker information during transmission, wherein the data encoding unit is driven to perform an encoding process on the system tuning information, residual-feedback data or training-breaker information, so as to reduce the risk of leaking sensitive information. Moreover, the state deletion unit is driven to perform a deletion process for required specific system state to avoid the leakage of the parameters related to important training processes in an event of system intrusion, so as to reduce the risk of leaking sensitive information. Besides, the visualized monitoring unit is driven to perform a visualized interface process on a specific monitor device for providing a system administrator a remote monitoring mechanism. Meanwhile, the data compressing unit is driven to perform a data compressing process on an output data to facilitate the transmission for system tuning information, residual-feedback data or training-breaker information between different modules through different devices, so as to enhance the efficiency of bandwidth utilization between different transmission devices. The history tracing unit is driven to perform a history tracing mechanism for data stream in transmission to enhance data security and avoid a large number of repeated decision-making processes, so that the system tuning module can find out abnormal transmission data via reviewing history logs. Combining these elements in an implementation of software, hardware, or a combination of software and hardware into a system tuning module for transmitting priority or sensitive data stream through network environment not only can spread the risks via deploying these units in different places for the neuro-computing sub-system and the residual-backpropagation sub-system, but also can provide effective maintenance and management mechanisms for decision-making data with commercial value, so as to make the present invention can be implemented as a neural network artificial intelligence decision-making core system with data security design.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 is a structural block diagram of a system according to an embodiment of the present invention.

FIG. 2 is a dataflow diagram showing main data flow related to an embodiment of the present invention.

FIG. 3 A is a schematic diagram of the system according to an embodiment of the present invention.

FIG. 3 B is a schematic diagram of the unsupervised neural-network interface module according to an embodiment of the present invention.

FIG. 3 C is a schematic diagram of the neuro-computing sub-system according to an embodiment of the present invention.

FIG. 3 D is a structural block diagram of a server-virtualization platform according to an embodiment of user scenario.

FIG. 4 is a structural block diagram of a system according to an embodiment of the present invention.

FIG. 5 A is a block diagram of an asymmetric-hidden-layers input module according to an embodiment of the present invention.

FIG. 5 B is a block diagram of an asymmetric-hidden-layers input module according to an embodiment of the present invention.

FIG. 5 C is a dataflow diagram showing main data flow related to a non-linear PCA computing process according to an embodiment of the present invention.

FIG. 5 D is a dataflow diagram showing main data flow related to a non-linear PCA computing process according to an embodiment of the present invention.

FIG. 5 E is a block diagram of a tree-structured neuron module according to an embodiment of the present invention.

FIG. 6 is a block diagram of a layered weight parameter module according to an embodiment of the present invention.

FIG. 7 is a block diagram of a tree-searching module according to an embodiment of the present invention.

FIG. 8 is a structural block diagram of a residual-backpropagation sub-system according to an embodiment of the present invention.

FIG. 9 is a block diagram of a system tuning module according to an embodiment of the present invention.

DETAILED DESCRIPTION OF THE INVENTION

To facilitate understanding of the objects, characteristics and effects of this present disclosure, embodiments together with the attached drawings for the detailed description of the present disclosure are provided.

A plurality of embodiments in regard to an artificial-intelligence decision-making core system with neural network which can be applied to a variety of trained data from external non-specific models (such as financial models, human behavior models, war chess models, etc.), to facilitate deriving decision-making data with applicability and individuality, are provided as follows. In an embodiment of this aspect, the system can be further implemented a decision-making core system with security protection mechanism. In some embodiments, this technology can be implemented as a combination of a disk-array management device, an embedded system and a server-virtualization platform, based on a Software Defined Network (SDN) infrastructure, and data transmission between sub-systems can be performed via network functions virtualization (NFV) components including virtual gateway/router for data transmission channels between different virtual network segments.

Refer to FIG. 1 , which is a structural block diagram of a decision-making core system according to an embodiment of the present invention. The artificial-intelligence decision-making core system with neural network S 1000 comprises an unsupervised neural-network interface module M 100 , a neuro-computing sub-system S 1001 , and a residual-backpropagation sub-system S 1002 . In an embodiment, the unsupervised neural-network interface module M 100 comprises at least a computing-and-storage device (or a computing device), wherein this device can be implemented as an input device for an unsupervised neural network to receive a raw data, and an interface of this device for receiving the raw data can be applied to a variety of trained data from external non-specific models (such as financial models, human behavior models, war chess models, etc.). A conventional neural-network input device for decision-making analysis is usually only applicable for raw data from specific model. For example, a neural-network device for health analysis usually cannot accept the history data of stock market through the same design. However, the interface of the present invention for receiving the raw data can be applied to various trained data from non-specific models via general classification settings (such as parameter configuration, attribute configuration, or labeling configuration, but not limited to the above configurations). In an embodiment, M 100 can be implemented as a server with a disk-array management unit, and a variety of trained data from external non-specific models can be uploaded to M 100 as the raw data of the system S 1000 . The disk-array management unit of M 100 makes use of a random-storage algorithm (such as adopting array or dynamic arrays co-working with a combination of merge-sorting operations, or other random-storage algorithm which is well known to a person having ordinary skill in the art related to the present invention) to distribute the raw data into disk array, and the physical storage media of the disk array is not limited to be installed in a single server. Through such interface management mechanism of the unsupervised neural-network interface module M 100 , this device not only can extract a specific trained data set quickly, but also provides a data risk management framework for data storage. In addition, this embodiment is only a feasible example, and the implementation of the module M 100 is not limited thereto.

In an embodiment, the above-mentioned neuro-computing sub-system S 1001 is implemented with at least one computing device integrated with or coupled to a storage device, connected to the unsupervised neural-network interface module M 100 , for performing a pre-processing computing process (for example, comprising operations of data cleaning, data editing, data reduction, data wrangling, and so on, such like some conventional data pre-processing routines for neural-network deep learning) on the raw data to obtain pre-processed data. Thereafter, S 1001 performs a neuron operation process that can dynamically update the neuron node, and then derives corresponding non-linear analyzed data and decision-making data, via a non-linear PCA computing process. A conventional neuron operation process is usually based on the following input/output relation: output=f activation (Î£(weight*input)+bias). However, in the present invention, the above-mentioned the âneuron operation process that can dynamically update the neuron nodeâ means not only weight and input but also bias, and even the activation function can be updated dynamically through this neuron operation process. In an embodiment, the sub-system S 1001 can be implemented as an embedded system integrated with a Neural Compute Stick (NCS), where the neural network according to the present invention can be implemented on this neural compute stick practically, and one or more processors of the embedded system can co-work with a specific combination of instructions stored in the embedded system to perform the above mentioned operation processes related to S 1001 .

Referring to FIG. 1 , in an embodiment, the above-mentioned residual-backpropagation sub-system S 1002 is coupled to the neuro-computing sub-system S 1001 , for performing a residual compensation computing process based on the non-linear analyzed data derived from S 1001 , to generate system tuning information and residual-feedback data, or training-breaker information, to facilitate the neuro-computing sub-system S 1001 to derive individual and applicable decision-making data. The system tuning information is used by the neuro-computing sub-system S 1001 to correct or tune its neural network. The training-breaker information is used by the neuro-computing sub-system S 1001 to perform a training-break mechanism on a training process in the neural network. The residual-feedback data is to facilitate the neuro-computing sub-system S 1001 to derive or correct the decision-making data. In an embodiment, the sub-system S 1002 can be implemented as a server-virtualization platform with a plurality of virtual machines (VMs), wherein one of the VMs can be performed as a monitor-and-management unit to coordinate the other VMs to complete the above-mentioned operation processes related to S 1002 . However, this embodiment is only a feasible example, and the implementation of the sub-system S 1002 is not limited thereto.

Refer to FIG. 2 , which is a dataflow diagram showing main data flow related to an embodiment of the artificial-intelligence decision-making core system with neural network S 1000 shown in FIG. 1 . As shown in FIG. 2 and also referring to FIG. 1 , a dataflow is illustrated based on one of user scenarios for implementing S 1000 of an embodiment of the present invention. First, as shown in block F 100 , the unsupervised neural-network interface module M 100 receives the trained data set from the data source as the raw data of the system. Then, as shown in block F 101 , the raw data is converted into an input vector which is applicable for an unsupervised neural network, whereby the raw data is imported into the input layer of an unsupervised neural network. After that, as shown in block F 102 , the raw data is imported to an asymmetric hidden layer from the input layer of the neural network, and a pre-processing process is performed through neural network operations. Then, as shown in block F 103 , the system S 1000 completes the pre-processing process of the asymmetric hidden layer, and output the pre-processed data through the output layer of the neural network. Thereafter, as shown in block F 110 , the system S 1000 performs a process for tree-like data structure according to the pre-processed data, to generate the tree-structured data corresponding to each neuron node, and whereby the neuron node can be dynamically updated; here the mentioned neuron node which can be dynamically updated is mainly located in the asymmetric hidden layer shown in block F 102 . Once the F 102 -mentioned asymmetric hidden layer changes the states of a neuron node during a training process, the F 110 -mentioned tree-structured data will be updated as well. Thereafter, as shown in block F 111 , the system S 1000 performs a layered weight computing process according to the tree-structured data corresponding to the neuron node, in the layered weight parameter module, to obtain a data of layered weight parameter, thereby manages each specific layered weight of the asymmetric hidden layers; here the F 111 -mentioned the data of layered weight parameter is for use in a training process related to the asymmetric hidden layers through a management mechanism. Thereafter, as shown in block F 112 , the data of layered weight parameter will be processed within the dataflow of block F 112 ; according to the data of layered weight parameter, a non-linear principal component analysis (PCA) computing process is performed to generate a non-linear analyzed data, whereby the applicable decision-making data can be derived, based on or coupled to, the residual-feedback data delivered from the residual-backpropagation sub-system. So far, the dataflow shown in the above blocks F 101 -F 112 can be implemented in the neuro-computing sub-system S 1001 as shown in FIG. 1 . However, in order to maintain the applicability and individuality of the derived decision-making data mentioned in block F 112 , the non-linear analyzed data mentioned in block F 112 will be introduced to the dataflow shown by block F 200 .

As shown in block F 200 , the residual-backpropagation sub-system performs a residual compensation computing process based on the non-linear analyzed data to generate residual-feedback data to the non-linear PCA module, or break the training processes based on the training-breaker information for the neural network, or deliver the system tuning information to facilitate correcting or tuning the neural network; here the block F 200 -mentioned residual-feedback data will be introduced back to the dataflow shown in block F 112 for correcting the decision-making data, and the block F 200 -mentioned training-breaker information or system tuning information will be introduced to the dataflow shown in block F 201 . As described in block F 201 , the residual-backpropagation sub-system sends the training-breaker information or the system tuning information to a tree-searching module, and couples the tree-structured neuron module to break the training processes of the neural network or to dynamically modify the neuron nodes. So far, the dataflow shown in the above block F 200 can be implemented in the residual-backpropagation sub-system S 1002 as shown in FIG. 1 , and the dataflow described in block F 201 goes from sub-system S 1002 to sub-system S 1001 ; wherein, the block F 201 -mentioned tree-searching module is located in the sub-system S 1001 , and this module is also the entry-point of the sub-system S 1001 for receiving the training-breaker information or the system tuning information derived from the sub-system S 1002 ; through a tree-searching process, here the tree-searching module can find an index of the neuron node which is required to be processed, and then the index will be delivered with the training-breaker information or the system tuning information to the block F 201 -mentioned tree-structured neuron module. Thereafter, the tree-structured neuron module will perform a training-break mechanism according to the index with the training-breaker information through the dataflow related to the block F 102 to break a training process; or perform to update neuron nodes dynamically according to the index with the system tuning information through the dataflow related to the block F 110 , so as to tune the neural network. Based on the above dataflow, the implementation of each component corresponding to the present invention will be described as follows.

Referring to FIG. 1 , in the above embodiment, the neuro-computing sub-system S 1001 comprises: an asymmetric-hidden-layers input module M 200 , a tree-structured neuron module M 201 , a layered weight parameter module M 202 , a non-linear PCA module M 203 , and a tree-searching module M 204 . The asymmetric-hidden-layers input module M 200 can perform a data pre-processing program on raw data sent from the unsupervised neural-network interface module M 100 of the artificial-intelligence decision-making core system S 1000 , to derive a pre-processed data; here the asymmetric-hidden-layers input module M 200 is formed by unsupervised neural network, and the asymmetric-hidden-layers means that the bidirectional weights between neuron nodes in the hidden layers are not equal. The tree-structured neuron module M 201 for managing neuron nodes of a neural network is connected to the asymmetric-hidden-layers input module M 200 , including a plurality of neuron-node data (or data representing neuron nodes), wherein the module can perform a tree-structured data processing program based on the received pre-processed data, to generate tree-structured data corresponding to each neuron node. Meanwhile, M 201 can correct the tree-structured data of neuron nodes according to the system tuning information derived from the residual-backpropagation sub-system S 1002 . The layered weight parameter module M 202 is connected to the tree-structured neuron module M 201 and performs a layered weight parameter computing process according to the tree-structured data derived from M 201 , to obtain a layered weight parameter data for managing the specific weight of each hidden layer in the neural network, so that in addition to parameters related to individual neurons, the layered weight factors should also be considered for tuning a neural network. For example, the above-mentioned layered weight parameter computing process comprises the steps of matrixing individual neuron weights included in each layer of the hidden layers, normalizing by geometric average method to extract the eigenvector, and vectorizing the layered weights according to the consistency index. However, the combination of the above steps is only a feasible example of the present invention, and its main purpose is to integrate the weight parameter data of the layered individual neurons to derive the vectorized layered weight relationship applicable to the raw data. In addition, the non-linear PCA (principal component analysis) module M 203 for deriving a non-linear analyzed data is connected to the layered weight parameter module M 202 to perform a non-linear PCA computing process according to the layered weight parameter data derived by M 202 , wherein the non-linear PCA module M 203 can derive applicable decision-making data based on the non-linear analyzed data or can work with the residual-feedback data from the residual-backpropagation sub-system S 1002 to improve the decision-making data. The tree-searching module M 204 is coupled to the tree-structured neuron module M 201 and the residual-backpropagation sub-system S 1002 for breaking the neural-network training process in time to maintain the differentiation and the applicability of the decision-making data.

For example, the modules (M 200 -M 204 ) mentioned in the above embodiments can be implemented with part or all of at least one storage device, and store combinations of specific instructions or prog

CLAIMS

Claims ( 12 )

What is claimed is:

1. An artificial-intelligence decision-making core system with neural network, the system comprising:

an unsupervised neural-network interface module which comprises at least a computing device for receiving raw data, wherein an interface of the unsupervised neural-network interface module for receiving the raw data is applicable for trained data from external non-specific models;

a neuro-computing sub-system which comprises at least a computing device coupled to the unsupervised neural-network interface module, wherein the neuro-computing sub-system is configured to perform a pre-processing computing process on the raw data for obtaining pre-processed data, and then to perform a neuron operation process capable of dynamically updating neuron node, and then to derive corresponding non-linear analyzed data and decision-making data, via a non-linear principle component analysis (PCA) computing process; and

a residual-backpropagation sub-system, in cooperation with the neuro-computing sub-system, for performing a residual compensation computing process based on the non-linear analyzed data to generate system tuning information and residual-feedback data, or training-breaker information, and to deliver the system tuning information and the residual-feedback data, or the training-breaker information to the neuro-computing sub-system, so as to facilitate the neuro-computing sub-system to derive decision-making data with individuality and applicability;

wherein the neuro-computing sub-system comprises:

an asymmetric-hidden-layers input module, for performing a data pre-processing program on the raw data sent from the unsupervised neural-network interface module of the artificial-intelligence decision-making core system, to derive pre-processed data;

a tree-structured neuron module, coupled to the asymmetric-hidden-layers input module, including a plurality of neuron-node data, wherein the tree-structured neuron module is configured to perform a tree-structured data processing program based on the received pre-processed data, so as to generate tree-structured data corresponding to each neuron node;

a layered weight parameter module, coupled to the tree-structured neuron module, for performing a layered weight parameter computing process according to the tree-structured data, so as to obtain layered weight parameter data for managing a specific weight of each hidden layer in neural network;

a non-linear PCA module, coupled to the layered weight parameter module, for performing a non-linear PCA computing process according to the layered weight parameter data for deriving non-linear analyzed data, wherein the non-linear PCA module is configured to derive applicable decision-making data based on the non-linear analyzed data or is configured to work with the residual-feedback data from the residual-backpropagation sub-system so as to improve the decision-making data; and

a tree-searching module, coupled to the tree-structured neuron module and the residual-backpropagation sub-system, for breaking a training process of the neural network at once to maintain the differentiation and the applicability of the decision-making data.

2. The artificial-intelligence decision-making core system with neural network according to claim 1 , wherein in order to enhance the decision-making quality of the core system, the neuro-computing sub-system further comprises:

a hidden-layered routing module for receiving the raw data from the unsupervised neural-network interface module of the decision-making core system, and planning a route for the hidden layers according to attributes of the raw data, so that the neuro-computing sub-system supports variable propagation paths for the hidden layers of the neural network dynamically;

a dense data processing module for analyzing the raw data from the unsupervised neural-network interface module of the decision-making core system and perform a dense data processing program for dropping invalid data if there is a high-density attribute with the raw data, and normalizing duplicate data, and then importing the output data which has completed the dense data processing program, into the asymmetric-hidden-layers input module of the neuro-computing sub-system;

a sparse tensor processing module, coupled to the tree-structured neuron module of the neuro-computing sub-system, for verifying whether derived neuron nodes fit conditions of sparse tensor, wherein when neurons of the neural network are in a condition of sparse tensor, the sparse tensor processing module performs a sparse matrix operation process to facilitate computing performance and save system memory usage;

a Chaos model analysis module, coupled to the non-linear PCA module of the neuro-computing sub-system, for performing a Chaos model analysis process to facilitate derivation of the decision-making data, when the non-linear PCA module cannot make effective corrections according to the residual-feedback data sent from the residual-backpropagation sub-system; and

a neuron dimension switching module, coupled to the tree-structured neuron module, for performing a neuron dimension switching process based on system tuning information from the residual-backpropagation sub-system, to facilitate switching of decision-making style, so that the neuron dimension switching module is capable of enhancing the differentiation and applicability of the decision-making data.

3. The artificial-intelligence decision-making core system with neural network according to claim 1 , wherein the residual-backpropagation sub-system comprises:

a progressive backpropagation input module for receiving non-linear analyzed data from the non-linear PCA module;

a system tuning module to perform a residual compensation computing process for the non-linear analyzed data, to derive feedback-tuning information, and to trigger a training-break mechanism of neural network according to the feedback-tuning information or deliver the feedback-tuning information; and

a progressive backpropagation output module, based on an updated state of the system tuning module and the derived feedback-tuning information, to derive residual-feedback data and send the residual-feedback data to the non-linear PCA module of the neuro-computing sub-system for facilitating the neuro-computing sub-system to maintain the differentiation and applicability of the decision-making data.

4. The artificial-intelligence decision-making core system with neural network according to claim 3 , wherein in order to enhance the processing ability for non-linear analyzed data, the residual-backpropagation sub-system further comprises:

a dynamic-norm processing module for non-linear analyzed data that is capable of being vectorized or matrixed, wherein the progressive backpropagation input module of the residual-backpropagation sub-system is configured to further coordinate the dynamic-norm processing module to perform a pre-processing program to import the vectorized data or matrixed data to the system tuning module for operation of residual-compensation mechanism;

wherein if the non-linear analyzed data reveals that most of the layered weights are in a large scale, the progressive backpropagation input module of the residual-backpropagation sub-system further coordinates the dynamic-norm processing module to perform a re-initialization process to implement a norm-penalty mechanism for hidden layers to re-initialize related weight parameters of the neural network for reducing possible overfitting;

an inner-product processing module for vectorized data or matrixed data with image properties, wherein the progressive backpropagation input module of the residual-backpropagation sub-system is further configured to coordinate the inner-product processing module to perform another pre-processing program to self-define the relationship between nodes and vectors, so as to introduce standard orthogonal basis to the system tuning module to improve the operation of residual-compensation mechanism;

a manifold operation module for non-Euclidean space data, wherein the progressive backpropagation input module of the residual-backpropagation sub-system is configured to further coordinate the manifold operation module to perform another pre-processing program to switch the non-Euclidean space data into a form of Euclidean space, so as to import sampling data from various geometry spaces to the system tuning module to enhance the applicability of residual-compensation mechanism;

a tensor switching module for time-variant data, wherein the system tuning module of the residual-backpropagation sub-system is configured to further coordinate the tensor switching module to perform a self-correction process to switch the residual data already processed by residual-compensation mechanism through various dimensional spaces, so as to find the optimal residual-operation space that is most feasible for the time-variant data to enhance the applicability of residual-compensation mechanism;

a functional operation module for a computing process of obtaining the optimal solution of loss function via making use of boosting algorithm, wherein the system tuning module of the residual-backpropagation sub-system is configured to further coordinate the functional operation module to perform a mathematical programming process with functional operations, to facilitate derivation of the optimal solution of the loss function, so as to maintain accuracy of the residual-compensation mechanism; and

an ARIMA-model processing module for non-stationary residual data, wherein the system tuning module of the residual-backpropagation sub-system is configured to further coordinate the ARIMA-model processing module to perform a stationary process through integrated finite-difference operations to obtain stationary residual data, so as to check whether the stationary residual data fits a white-noise sequence to confirm whether original non-stationary residual data needs to be corrected.

5. The artificial-intelligence decision-making core system with neural network according to claim 1 or 3 , wherein the tree-structured neuron module of the neuro-computing sub-system further comprises:

a cross-neuron computing unit for performing a neuron data updating process according to the residual-feedback data received by the neuro-computing sub-system from the progressive backpropagation output module of the residual-backpropagation sub-system, so as to generate corresponding neuron updating data; and

a neuron-weight updating unit for performing another neuron data updating process according to the neuron updating data, so as to update the tree-structured data and the weight parameter data for neurons of the neural network, and notify the asymmetric-hidden-layers input module to perform an updating process for the raw data in a pre-processing process.

6. The artificial-intelligence decision-making core system with neural network according to claim 2 or 4 , wherein in order to prevent being hacked or learned via external neural networks, the tree-structured neuron module of the neuro-computing sub-system further comprises:

a neuron generator unit for performing a neuron generating process according to instructions from the tree-structured neuron module, so as to generate the required nodes for corresponding neurons;

a neuron topology switching unit for performing a neuron topology switching process according to the instructions from the tree-structured neuron module, working with the neuron generator unit, so as to dynamically change the topology structure of the neural network; and

a neuron grouping management unit for performing a neural-network reorganizing process for neuron nodes with various attributes, according to the instructions from the tree-structured neuron module, working with the neuron topology switching unit, so as to dynamically combine multiple neural networks with different topology structures.

7. The artificial-intelligence decision-making core system with neural network according to claim 1 , wherein the layered weight parameter module of the neuro-computing sub-system further comprises:

an activation-function integration unit for working with various activation functions to perform a setting process according to configured data of activation function managed by the layered weight parameter module and come out a setting result, so as to set an applicable activation function to a layered-weight parameter updating process for the neural network;

a Laplace-transformation processing unit for performing a Laplace transformation process according to the activation function configured by the layered weight parameter module and the weight-parameter data of the neurons to generate a Laplace operation result; and

an activation-function updating unit for performing an activation-function updating process to derive a configured data of activation function corresponding to the layered-weight parameter data according to the Laplace operation result, wherein the layered weight parameter module is configured to analyze and evaluate whether to coordinate the activation-function integration unit to reconfigure the corresponding activation function.

8. The artificial-intelligence decision-making core system with neural network according to claim 7 , wherein in order to enhance the generalization ability of the neural network, the layered weight parameter module of the neuro-computing sub-system further comprises:

a normalization processing unit for optimizing performance of hidden layers, wherein the layered weight parameter module is configured to drive the normalization processing unit according to the setting result from the activation-function integration unit, to perform a normalization processing program for an unbounded activation function;

a harmonic-function processing unit for the raw data that has phasing properties or is expressed on a complex plane, wherein the layered weight parameter module is configured to drive the harmonic-function processing unit according to the setting result from the activation-function integration unit, to perform a harmonic-function mapping process, so that the configured activation function is capable of supporting data which is in a form of complex number;

a time-variant function processing unit configured to be driven to import time variables to the configured activation function if the layered weight parameter module determines that the hidden layer is unable to avoid gradient vanishing problem or gradient exploding problem according to the configured data of activation function, so as to facilitate a result of weight adjustment tending to converge;

a multi-dimensional moment processing unit configured to be driven to introduce an affine transformation process for a moment generating function to the configured activation function according to the configured data of activation function if the layered weight parameter module determines that the raw data fits the property of multi-variant continuous random variables, so as to facilitate the result of weight adjustment tending to converge;

a Fourier-transformation processing unit configured to be driven to perform a Fourier-transformation process on the data propagated by the hidden layers via making use of the Dirac delta function to obtain a transformed data if the layered weight parameter module determines that the raw data has the distribution property of the Dirac delta function according to the configured data of activation function, thereby the Fourier-transformation processing unit uses a ramp function as an activation function to import the transformed data into the activation function, for supporting the raw data with particle properties or the raw data in zero dimension; and

a wave-function processing unit configured to be driven to perform a wave-function processing program to use an applicable wave function as the configured activation function according to the configured data of activation function if the layered weight parameter module determines that the raw data has distribution property of a quantum many-body system, so as to facilitate the result of weight adjustment for quantum many-body system simulation tending to converge.

9. The artificial-intelligence decision-making core system with neural network according to claim 1 or 3 , wherein the tree-searching module of the neuro-computing sub-system further comprises:

a duplicate-neurons preventing unit for determining whether there are duplicate neuron nodes according to the tree-structured data derived by the tree-structured neuron module of neuro-computing sub-system, wherein the duplicate-neurons preventing unit derives a duplicated-neuron information if there are duplicate neurons, so that the tree-searching module is capable of driving a state register unit of the residual-backpropagation sub-system to derive a corresponding training-breaker information, so that the neuro-computing sub-system coordinates the tree-structured neuron module of the neuro-computing sub-system according to the training-breaker information, to perform a training-break mechanism; and

an upper confidence bounds applied to trees (UCT) searching unit for performing an upper confidence bound (UCB) computing process based on the system tuning information from the residual-backpropagation sub-system and the raw data, tree-structured data, weight-parameter data, and non-linear analyzed data of the neuro-computing sub-system, to generate a corresponding upper-confidence-bound information, wherein the tree-searching module determines whether the current neural network does not fit the UCB information and coordinate the state register unit to derive a training-breaker information if it does not, so that the neuro-computing sub-system is enabled to coordinate the tree-structured neuron module of the sub-system according to the training-breaker information, to perform a training-break mechanism.

10. The artificial-intelligence decision-making core system with neural network according to claim 2 or 4 , wherein in order to enhance the tree-searching performance, the tree-searching module of the neuro-computing sub-system further comprises:

a support vector machine (SVM) processing unit configured to be driven to perform a hyperplane optimization process if the tree-searching module determines the current upper-confidence-bound is out of an applicable range according to the upper-confidence-bound information from the UCT searching unit, so as to adjust the decision-making boundary and optimize the upper-confidence-bound for avoiding too many invalid searching processes;

a genetic algorithm (GA) searching unit configured to be driven to perform a genetic-algorithm computing process for a neural network composed of a variety of topological structures, so as to group neuron nodes of the neural network to determine a target group, thereafter to perform a tree-structured data searching process on the target group, so that the tree-searching module can effectively simplify the searching scope;

a Monte Carlo tree searching unit configured to be driven to perform an MCTS process for a neural network composed of a tree-structured topology with a large depth, so that the tree-searching module is enabled to enhance the tree-searching performance through heuristic-searching processes, for reducing cost of paths;

a searching-schedule accelerating unit configured to be driven to perform a B+tree searching process in a file system embedded with B+tree file structure, for a neural network composed of a tree-structured topology with a large width, so that the tree-searching module is enabled to perform the searching schedule with B+tree searching algorithm, which facilitates effective saving of the search time; and

a hash-table searching unit configured to be driven to store indexes in a hash table and is configured to perform an index-searching process with hash-searching algorithm, for a neural network composed of a tree-structured topology with a large number of indexes, so that the tree-searching module can effectively save search time.

11. The artificial-intelligence decision-making core system with neural network according to claim 1 or 3 , wherein the system tuning module of the residual-backpropagation sub-system further comprises:

a residual analysis unit, with more than one residual computing program, wherein the residual analysis unit is configured to perform an applicable residual computing program according to the non-linear analyzed data imported by the progressive backpropagation input module to generate a residual analysis data;

a self-tuning processing unit for performing a self-tuning processing program based on the non-linear analyzed data and the residual analysis data, with an adaptive residual analysis for a backpropagation of residual compensation, to derive a system tuning information, wherein the system tuning information is for fine-tuning the neural network or for triggering a training-break mechanism;

a state register unit wherein the system tuning module is configured to store more than one break-configuration switching data in the state register unit according to the system-state data; and

a training breaker unit for determining whether a training-break condition is met according to the system tuning information from the self-tuning processing unit, wherein the training breaker unit works with the state register unit if the training-break condition is met, to generate a training-breaker information according to a selected break-configuration switching data, so that the residual-backpropagation sub-system is configured to further coordinate the neuro-computing sub-system to perform a training-break mechanism according to the training-breaker information.

12. The artificial-intelligence decision-making core system with neural network according to claim 2 or 4 , wherein in order to enhance the management for training states and system tuning information, the system tuning module of the residual-backpropagation sub-system further comprises:

a state recovery unit for a system that has already performed a training-break mechanism, wherein if the factor that caused the training-break has been eliminated, the state recovery unit will be driven to facilitate the system to restore a normal state which is before the training-break mechanism is triggered, according to the break-configuration switching data stored in the state register unit;

a state switching unit configured to be driven to perform a state switching process according to the break-configuration switching data stored in the state register unit if the current decision-making result is more degraded than the previous decision-making result or there are multiple types of raw data that must be switched, so as to facilitate the system to switch to a previous state which is more applicable;

a data encoding unit which is optional to couple a data compressing unit, for enhancing the confidentiality of the data delivered to the neuro-computing sub-system, wherein the data encoding unit is driven by the residual-backpropagation sub-system to perform an encoding process on the data delivered to the neuro-computing sub-system such as system tuning information, residual-feedback data or training-breaker information, so as to reduce the risk of leaking sensitive information in public network environment during the data transmission cross sub-systems;

a state deletion unit configured to be driven to perform a deletion process for required specific system state to avoid the leakage of the parameters related to important training states in an event of system intrusion, so as to reduce the risk of leaking sensitive information;

a visualized monitoring unit configured to be driven to perform a visualized interface process on a specific monitor device for providing a system administrator a remote monitoring mechanism;

a data compressing unit configured to be driven to perform a data compressing process on an output data to facilitate the transmission for system tuning information, residual-feedback data or training-breaker information between different modules through different devices, so as to enhance efficiency of bandwidth utilization between different transmission devices; and

a history tracing unit configured to be driven to perform a history tracing mechanism for data stream in transmission to enhance data security and avoid a large number of repeated decision-making processes, so that the system tuning module is enabled to find out abnormal transmission data via reviewing history logs.

US16/932,863

2020-07-02

2020-07-20

Artificial-intelligence decision-making core system with neural network

Active

2041-07-08

US11526726B2

( en )

Applications Claiming Priority (2)

Application Number

Priority Date

Filing Date

Title

TW109122440

2020-07-02

TW109122440A

TWI746038B

( en )

2020-07-02

Neural network-like artificial intelligence decision-making core system

Publications (2)

Publication Number

Publication Date

US20220004839A1

US20220004839A1 ( en )

2022-01-06

US11526726B2

true

US11526726B2 ( en )

2022-12-13

Family

ID=79167461

Family Applications (1)

Application Number

Title

Priority Date

Filing Date

US16/932,863

Active

2041-07-08

US11526726B2

( en )

2020-07-02

2020-07-20

Artificial-intelligence decision-making core system with neural network

Country Status (2)

Country

Link

( 1 )

US11526726B2

( en )

( 1 )

TWI746038B

( en )

Cited By (2)

* Cited by examiner, â Cited by third party

Publication number

Priority date

Publication date

Assignee

Title

US20230111546A1

( en )

2021-10-13

2023-04-13

Sigmastar Technology Ltd.

Image processing device and image processing method

US12015585B2

( en )

2022-04-29

2024-06-18

Bank Of America Corporation

System and method for detection, translation, and categorization of visual content associated with malicious electronic communication

Families Citing this family (41)

* Cited by examiner, â Cited by third party

Publication number

Priority date

Publication date

Assignee

Title

CN111709252B

( en )

2020-06-17

2023-03-28

åäº¬ç¾åº¦ç½è®¯ç§ææéå¬å¸

Model improvement method and device based on pre-trained semantic model

EP4229511A1

( en )

2020-10-16

2023-08-23

Neuroblade, Ltd.

Memory appliances for memory intensive operations

DE102022125014A1

( en )

2021-09-29

2023-03-30

Kinaxis Inc.

Methods and systems of sequential latent inference of complex knowledge

CN114492740B

( en )

2022-01-07

2025-06-03

è¥¿åå·¥ä¸å¤§å¦

A neural network model testing method based on weighted neuron coverage

CN114584358B

( en )

2022-02-25

2023-10-13

å®æ·åéç§ææé½æéå¬å¸

Intelligent network security system, device and storage medium based on Bayesian regularization

CN114580890B

( en )

2022-03-02

2024-07-26

åé·é¡¿ç§æè¡ä»½å¬å¸

Collaborative decision analysis neural network algorithm system design method, system and vehicle

CN114545884B

( en )

2022-03-16

2023-12-05

æ¸©å·å¤§å¦

Equivalent parallel machine dynamic intelligent scheduling method based on enhanced topological neural evolution

CN114626635B

( en )

2022-04-02

2024-08-06

åäº¬ä¹æºç§ææéå¬å¸

A steel logistics cost prediction method and system based on hybrid neural network

TWI819578B

( en )

2022-04-20

2023-10-21

åç«ä¸å¤®å¤§å¸

Multi-objective parameters optimization system, method and computer program product thereof

CN115033903B

( en )

2022-06-13

2024-07-02

æ¹åå·¥ä¸å¤§å¦

A privacy protection method and system for incomplete data in neural network training

CN115002703B

( en )

2022-07-14

2025-03-21

è¥¿å®çå·¥å¤§å¦

A passive indoor occupancy detection method based on Wi-Fi channel status information

TWI805511B

( en )

2022-10-18

2023-06-11

åç«ä¸æ£å¤§å¸

Device for computing an inner product

CN115604011B

( en )

2022-10-19

2023-08-04

ä¸æµ·èé«ç§ææéå¬å¸

OSI protocol multi-point communication method based on robot and block chain

CN115829016B

( en )

2022-12-15

2024-01-16

å¹¿å·é²é¦éç©èç½ç§æè¡ä»½æéå¬å¸

Adjustment method and device for wireless modem based on neural network

CN116306770B

( en )

2023-02-17

2023-11-14

è¥¿å®çµåç§æå¤§å¦å¹¿å·ç ç©¶é¢

Software defined network performance prediction method based on dense parent neural architecture search

CN116088579A

( en )

2023-02-27

2023-05-09

éåºé®çµå¤§å¦

Unmanned aerial vehicle self-adaptive information path planning method based on deep reinforcement learning

CN116701530B

( en )

2023-04-25

2025-09-26

æé½è¶ç®ä¸å¿è¿è¥ç®¡çæéå¬å¸

A supercomputing data center system and method

CN116705057B

( en )

2023-05-30

2026-04-07

å®æ³¢ä¸æµ·éå¢æéå¬å¸

A method, system, storage medium, and intelligent terminal for detecting leaks in water supply pipelines.

CN116719293B

( en )

2023-08-04

2023-11-17

æ±èç¬ééå±ç§ææéå¬å¸

Production monitoring method and system for cold-rolled steel strip

CN117075549B

( en )

2023-08-17

2024-03-22

æ¹åæºè¾¾æºè½ç§ææéå¬å¸

Plant control method and system based on artificial neural network

CN116821774B

( en )

2023-08-28

2023-11-28

å±±ä¸çå·¥èä¸å¦é¢

Power generation fault diagnosis method based on artificial intelligence

CN117035125B

( en )

2023-10-09

2024-01-02

ä¸çµç§å¤§æ°æ®ç ç©¶é¢æéå¬å¸

A decision-making system, method and device for distributed artificial intelligence model

CN117274059B

( en )

2023-10-11

2025-07-01

æå·ä¸éåè±¡ç§ææéå¬å¸

Low-resolution image reconstruction method and system based on image coding-decoding

CN117236650B

( en )

2023-11-13

2024-03-01

å±±ä¸å·¥æ³µçµæºæéå¬å¸

Water-fertilizer integrated intelligent pump house control method

CN118378912B

( en )

2024-04-29

2024-11-15

æ·±å³å¸éææºèç§ææéå¬å¸

Emergency scene intelligent analysis and decision support method based on AI large model

US20250355823A1

( en )

2024-05-17

2025-11-20

Rosemount Inc

High speed communication field device platform

CN118277745B

( en )

2024-06-03

2024-08-06

éå²åä¸å¤§å¦

A visual agricultural big data analysis interactive system and method

GB2700827A

( en )

2024-06-21

2026-03-18

Standard Chartered Bank Singapore Branch

System and method for estimating confidence and implementing metacognitive abilities in artificial intelligence systems

CN118468934B

( en )

2024-06-25

2025-03-21

ä¸å½ç§å¦é¢åè¥ç©è´¨ç§å¦ç ç©¶é¢

A method for constructing a fusion neural network model for negative ion source performance prediction

CN118709724B

( en )

2024-08-29

2024-11-08

åäº¬èªç©ºèªå¤©å¤§å¦

Reduced manifold neural operator construction method for space-time process unequal domain mapping

CN119276910B

( en )

2024-12-09

2025-02-25

åå·è§æ³ç§æè¡ä»½æéå¬å¸

Equipment operation maintenance monitoring system based on Internet of things

CN119376739B

( en )

2024-12-30

2025-04-01

æå·æµ·åº·å¨è§æ°åææ¯è¡ä»½æéå¬å¸

Model optimization method and device, electronic equipment and storage medium

CN119847057B

( en )

2025-03-21

2025-10-28

ç¾å®æ¶é²ç§ææéå¬å¸

Energy storage and fire control intelligent linkage method and system based on hybrid control

CN119961887B

( en )

2025-04-09

2025-08-12

ä¸èå¸ä½å°çæºè½æºçµè¡ä»½æéå¬å¸

Variable frequency control method and system for ventilation system of nuclear power station

CN119961577B

( en )

2025-04-09

2025-07-15

åäº¬å¥¥ç»´äºç½å¤§æ°æ®ç§æè¡ä»½æéå¬å¸

Automatic AI corpus data screening and obtaining method based on big data analysis

CN120075927B

( en )

2025-04-21

2025-07-04

å®æ³¢æ±ååæ°å¹¿çµç½ç»æéå¬å¸

Three-in-one terminal equipment heterogeneous network fusion and intelligent switching control method and system

CN120301042B

( en )

2025-06-09

2025-08-19

æµæ±çé®çµå·¥ç¨å»ºè®¾æéå¬å¸

A control method and system for a multi-source intelligent power manager based on 5G communication

CN120354765B

( en )

2025-06-26

2025-08-22

æµæ±æºçµèä¸ææ¯å¤§å¦

Dynamic automatic analysis method of worm gear meshing center distance based on servo drive

CN120896968B

( en )

2025-09-09

2026-04-17

å¹¿å·æ²³ä¸ç§ææéå¬å¸

Internet of Things-based air quality sensors and their remote control methods

CN120834649B

( en )

2025-09-18

2025-12-09

ä¸æµ·çµåå¤§å¦

Active power distribution network real-time voltage control method based on deep learning

CN121863691A

( en )

2026-03-16

2026-04-14

åèå¤çµå(éå¢)æéè´£ä»»å¬å¸åå¤´ä¾çµåå¬å¸

A power distribution dispatching system and method based on dynamic load forecasting

Citations (5)

* Cited by examiner, â Cited by third party

Publication number

Priority date

Publication date

Assignee

Title

US7366564B2

( en )

2002-08-23

2008-04-29

The United States Of America As Represented By The Secretary Of The Navy

Nonlinear blind demixing of single pixel underlying radiation sources and digital spectrum local thermometer

US20140012789A1

( en )

2012-04-30

2014-01-09

The Regents Of The University Of California

Problem solving by plastic neuronal networks

US20170109628A1

( en )

2015-10-20

2017-04-20

International Business Machines Corporation

Resistive processing unit

US9646243B1

( en )

2016-09-12

2017-05-09

International Business Machines Corporation

Convolutional neural networks using resistive processing unit array

US9715656B1

( en )

2016-09-12

2017-07-25

International Business Machines Corporation

Killing asymmetric resistive processing units for neural network training

Family Cites Families (4)

* Cited by examiner, â Cited by third party

Publication number

Priority date

Publication date

Assignee

Title

US10409165B2

( en )

2014-12-15

2019-09-10

Asml Netherlands B.V.

Optimization based on machine learning

KR101880901B1

( en )

2017-08-09

2018-07-23

ííìíë¦¬í°ìì¤í ì£¼ìíì¬

Method and apparatus for machine learning

CN108171329A

( en )

2017-12-13

2018-06-15

ååå¸èå¤§å¦

Deep learning neural network training method, number of plies adjusting apparatus and robot system

CN110807520A

( en )

2020-01-08

2020-02-18

æé½åæ¹ä¼ä¸è½¯ä»¶è¡ä»½æéå¬å¸

Method and device for analyzing influence factors of neural nodes of convolutional neural network

2020

2020-07-02

TW109122440A

patent/TWI746038B/en

active

2020-07-20

US16/932,863

patent/US11526726B2/en

active

Active

Patent Citations (5)

* Cited by examiner, â Cited by third party

Publication number

Priority date

Publication date

Assignee

Title

US7366564B2

( en )

2002-08-23

2008-04-29

The United States Of America As Represented By The Secretary Of The Navy

Nonlinear blind demixing of single pixel underlying radiation sources and digital spectrum local thermometer

US20140012789A1

( en )

2012-04-30

2014-01-09

The Regents Of The University Of California

Problem solving by plastic neuronal networks

US20170109628A1

( en )

2015-10-20

2017-04-20

International Business Machines Corporation

Resistive processing unit

US9646243B1

( en )

2016-09-12

2017-05-09

International Business Machines Corporation

Convolutional neural networks using resistive processing unit array

US9715656B1

( en )

2016-09-12

2017-07-25

International Business Machines Corporation

Killing asymmetric resistive processing units for neural network training

Cited By (3)

* Cited by examiner, â Cited by third party

Publication number

Priority date

Publication date

Assignee

Title

US20230111546A1

( en )

2021-10-13

2023-04-13

Sigmastar Technology Ltd.

Image processing device and image processing method

US12293490B2

( en )

2021-10-13

2025-05-06

Sigmastar Technology Ltd.

Image processing device and image processing method using three-dimensional and artificial intelligence noise reduction

US12015585B2

( en )

2022-04-29

2024-06-18

Bank Of America Corporation

System and method for detection, translation, and categorization of visual content associated with malicious electronic communication

Also Published As

Publication number

Publication date

US20220004839A1

( en )

2022-01-06

TWI746038B

( en )

2021-11-11

TW202203093A

( en )

2022-01-16

Artificial-intelligence decision-making core system with neural network — Ahp-Tech Inc. (US11526726B2)

Related documents