Method for an explainable autoencoder and an explainable generative adversarial network
An exemplary embodiment provides an autoencoder which is explainable. An exemplary autoencoder may explain the degree to which each feature of the input attributed to the output of the system, which may be a compressed data representation. An exemplary embodiment may be used for classification, such as anomaly detection, as well as other scenarios where an autoencoder is input to another machine learning system or when an autoencoder is a component in an end-to-end deep learning architecture. An exemplary embodiment provides an explainable generative adversarial network that adds explainable generation, simulation and discrimination capabilities. The underlying architecture of an exemplary embodiment may be based on an explainable or interpretable neural network, allowing the underlying architecture to be a fully explainable white-box machine learning system.
1 . A non-transitory computer-readable medium comprising computer program code that, when executed by a computer comprising a processor and a memory, is configured to configure the computer to perform steps for providing an explainable model comprising one or more of an explainable autoencoder and an explainable generative adversarial network, the steps comprising:
forming a first model and a second model; wherein one or more of the first model and second model is a white-box explainable model comprising a first feature importance vector having a global level of explainability; and
generating, with the first model and the second model, simultaneously in a single feed-forward step, one or more rules, a plurality of feature attributions, and one or more explanations, by steps comprising:
generating synthetic data from the first model and inputting the synthetic data and/or an input dataset to the second model, and forming the explainable model, the explainable model comprising one or more of the explainable autoencoder and the explainable generative adversarial network based on an output of the second model associated with the synthetic data and/or the input dataset, wherein forming the explainable model comprises identifying a plurality of partitions associated with the synthetic data and/or the input dataset, wherein the partitions are arranged in a hierarchy, each partition in the hierarchy having a predetermined proximity to each other partition in the hierarchy, said predetermined proximity defined by the hierarchy, wherein each partition in the hierarchy comprises a second feature importance vector having an intermediate level of explainability;
extracting partition information from the explainable model and identifying the plurality of feature attributions associated with a plurality of features of the input dataset; and
generating the one or more explanations based on the partition information and the plurality of feature attributions; and
terminating processing associated with generating the one or more rules, the plurality of feature attributions, and the one or more explanations, and outputting the one or more explanations, following the single feed-forward step.
2 . The non-transitory computer-readable medium according to claim 1 , wherein the input dataset is input into the second model, wherein the input dataset comprises an output of a deep learning system.
3 . The non-transitory computer-readable medium according to claim 2 , wherein the deep learning system is one of a convolutional neural network (CNN) or an explainable convolutional neural network (CNN-XNN), and wherein the output of the deep learning system is an image.
4 . The non-transitory computer-readable medium according to claim 3 , wherein the image is an X-ray image, and wherein identifying the plurality of feature attributions associated with a plurality of features of the input dataset comprises highlighting one or more pixels in the image.
5 . The non-transitory computer-readable medium according to claim 4 , further comprising identifying a reconstruction loss in the X-ray image, comparing the reconstruction loss to a threshold, and identifying an anomaly based on comparison of the reconstruction loss to the threshold.
6 . The non-transitory computer-readable medium according to claim 1 , wherein the input dataset is a plurality of partial medical readings, and wherein the steps comprise applying at least one network reversibility method to reconstruct a medical image from the plurality of partial medical readings.
7 . The non-transitory computer-readable medium according to claim 1 , wherein the input dataset comprises a medical image and at least one additional feature, and wherein the steps comprise encoding the medical image into a latent space, and, after encoding the medical image into the latent space, connecting the at least one additional feature.
8 . The non-transitory computer-readable medium according to claim 7 , wherein the at least one additional feature is human knowledge and wherein the steps further comprise injecting human knowledge into the explainable model, wherein the human knowledge creates or alters the one or more rules or one or more of the plurality of partitions, and wherein the created or altered one or more rules or one or more of the plurality of partitions is static and unchangeable by the explainable model, and wherein the human knowledge is expressed as a knowledge graph comprising nodes that describe human defined entities, and edges that describe a relationship between nodes; and wherein each node in the knowledge graph comprises one or more taxonomy identifiers that contribute to a classification of knowledge in the knowledge graph, wherein the knowledge graph comprises multiple taxonomies; and wherein one or more taxonomy identifiers are configured to fuse a plurality of knowledge graphs together.
9 . The non-transitory computer-readable medium according to claim 1 , wherein the input dataset comprises a video comprising a plurality of images arranged in a sequence, and wherein identifying the plurality of feature attributions associated with a plurality of features of the input dataset comprises highlighting one or more pixels in each image in the sequence.
10 . The non-transitory computer-readable medium according to claim 1 , wherein the input dataset comprises a loan application, wherein the one or more explanations comprise attributes of the loan application contributing to acceptance or rejection, and wherein generating the one or more explanations comprises determining that at least one attribute of the loan application is associated with an illegal outcome state.
11 . The non-transitory computer-readable medium according to claim 1 , wherein the explainable model is configured to be used as the basis or part of a practical data privacy preserving artificial intelligence (AI) system implementation, including at least one of: a differential privacy solution, a secure multi-party computation solution, a federated learning solution, and a homomorphic encryption solution.
12 . The non-transitory computer-readable medium according to claim 11 , wherein the explainable model includes the differential privacy solution, comprising a step of introducing noise in training data governed by a predetermined noise factor.
13 . The non-transitory computer-readable medium according to claim 11 , wherein the explainable model includes the secure multi-party computation solution, comprising a step of performing a portion of computation steps in a protected space.
14 . The non-transitory computer-readable medium according to claim 11 , wherein the explainable model includes the federated learning solution, comprising a step of executing the explainable model across a plurality of devices, each holding local data samples not shared with each other device in the plurality of devices.
15 . The non-transitory computer-readable medium according to claim 11 , wherein the explainable model includes the homomorphic encryption solution, comprising a step of executing one or more computations directly on encrypted data.
16 . The non-transitory computer-readable medium according to claim 1 , further comprising a step of sending the one or more explanations to a workflow system, and receiving feedback data from the workflow system and updating the explainable autoencoder based on the feedback data.
17 . The non-transitory computer-readable medium according to claim 16 , wherein the workflow system is a component of a Robotic Process Automation (RPA) system.
18 . The non-transitory computer-readable medium according to claim 17 , wherein the input dataset comprises sequence data comprising sensing data including at least one of LIDAR data, RADAR data, SONAR data, or sensor data.
19 . The non-transitory computer-readable medium according to claim 16 , wherein the workflow system is a component of a Decision Support System (DSS).
20 . A computer apparatus comprising a processor and a memory, the memory comprising program code configured to configure the processor to perform steps for providing an explainable model comprising one or more of an explainable autoencoder and an explainable generative adversarial network, the steps comprising:
forming a first model and a second model; wherein one or more of the first model and second model is a white-box explainable model comprising a first feature importance vector having a global level of explainability; and
generating, with the first model and the second model, simultaneously in a single feed-forward step, one or more rules, a plurality of feature attributions, and one or more explanations, by steps comprising:
generating synthetic data from the first model and inputting the synthetic data and/or an input dataset to the second model, and forming the explainable model, the explainable model comprising one or more of the explainable autoencoder and the explainable generative adversarial network based on an output of the second model associated with the synthetic data and/or the input dataset, wherein forming the explainable model comprises identifying a plurality of partitions associated with the synthetic data and/or the input dataset, wherein the partitions are arranged in a hierarchy, each partition in the hierarchy having a predetermined proximity to each other partition in the hierarchy, said predetermined proximity defined by the hierarchy, wherein each partition in the hierarchy comprises a second feature importance vector having an intermediate level of explainability;
extracting partition information from the explainable model and identifying the plurality of feature attributions associated with a plurality of features of the input dataset; and
generating the one or more explanations based on the partition information and the plurality of feature attributions; and
terminating processing associated with generating the one or more rules, the plurality of feature attributions, and the one or more explanations, and outputting the one or more explanations, following the single feed-forward step.