JP-2022537250-A5 -

Dates

Publication Date

20230515

Application Date

20200506

Description

In the above specification, embodiments of the present invention have been described with reference to numerous specific details which may differ between embodiments. Accordingly, the sole and exclusive indicator of the claims and the claims intended by the applicant are the set of claims relating to this application, including any subsequent modifications, which are specifically described in the description of the claims. Accordingly, any limitations, elements, characteristics, features, advantages, or attributes not expressly described in the claims should not in any way limit the scope of those claims. Accordingly, the specification and drawings should be understood as illustrative, not restrictive. Preferred embodiments of the present invention are described below in separate sections. Embodiment 1 A computer implementation method for constraining the behavior of a neural network, The neural network is made to perform a first inference operation based on the first training data to generate a first output. Based on the first target output associated with the first training data, generate a first rule corresponding to the first output. Determining that the neural network generates the first output when performing the second inference operation, and Executing the first rule which prevents the neural network from outputting the first output in response to the reception of the first input, and causes the neural network to output the first target output in response to the reception of the first input, The method, including the method described above. Embodiment 2 The computer implementation method according to Embodiment 1, further comprising determining that the first output is inaccurate based on a comparison between the first output and the first target output. Embodiment 3 To generate a graphical user interface to display the program code associated with the first rule, To generate a second rule, receive at least one change to the program code via the graphical user interface, and Preventing the neural network from outputting the first output in response to receiving the first input, and executing the second rule such that the neural network outputs a different output in response to receiving the first input, The computer implementation method according to Embodiment 1, further comprising: Embodiment 4 Extracting a set of vocabulary terms related to the first training data from the knowledge base, and To generate a graphical user interface for displaying the aforementioned set of vocabulary terms, The computer implementation method according to Embodiment 1, further comprising: Embodiment 5 Extracting a set of domain facts from the knowledge base that represent one or more logical attributes of the first training data, and To generate a graphical user interface for displaying the aforementioned set of domain facts, The computer implementation method according to Embodiment 1, further comprising: Embodiment 6 The method further includes generating derived facts based on the set of domain facts and the first output, The computer implementation method according to Embodiment 5, wherein the aforementioned derived facts characterize the functional attributes of the neural network when performing the first inference operation. Embodiment 7 To generate an architectural representation of the neural network for display via a graphical user interface, and Based on the input received via the graphical user interface, generate multiple different versions of the neural network. The computer implementation method according to Embodiment 1, further comprising: Embodiment 8 To generate performance data for at least one version of the neural network, wherein the performance data characterizes one or more performance characteristics of the at least one version of the neural network during the training phase, and To update the graphical user interface in order to display the performance data, The computer implementation method according to Embodiment 1, further comprising: Embodiment 9 The computer implementation method according to Embodiment 8, wherein the performance data indicates the accuracy with which at least one version of the neural network generates one or more outputs during operation. Embodiment 10 The computer implementation method according to Embodiment 8, wherein the performance data indicates the length of time required for at least one version of the neural network to produce one or more outputs. Embodiment 11 A non-temporary computer-readable medium that stores program instructions for restricting the behavior of a neural network when executed by a processor, wherein the processor A step of generating a first rule corresponding to a first output based on a first target output associated with first training data, wherein the first output results from a first inference operation performed by the neural network using the first training data. A step of determining that the neural network generates the first