Abstract: Methods and systems for detecting attempted manipulation of a machine learning model via explanation poisoning are provided. The method includes: computing explanations based on outputs of the model that include information that relates to features that affect the output of the model with respect to the first data point; assigning labels to the explanations based on the features; generating an explanation ensemble that resides in an N-dimensional space, N being equal to a number of assigned labels plus one; determining a region within the N-dimensional space for which a subsequent introduction of data causes a subsequent explanation that does not relate to the features; and when the additional data is introduced to the determined region, generating an alert message for notifying a user that a likelihood of adverse manipulation of the model is high based on the additional data.

Abstract: A method for facilitating model refinement via evolutive computation for a predetermined characteristic is disclosed. The method includes capturing a model explanation for a model based on a sampled data set, the model explanation including a listing of model features based on feature scores; identifying feature sets based on the model explanation, the feature sets including a selection of the model features based on the listing; computing, by using raw data, projected data sets for each of the feature sets based on the corresponding selection; generating a data set graph based on the projected data sets, the data set graph representing a relationship between each of the projected data sets; selecting, by using the data set graph, the projected data sets; and validating the selected projected data sets based on a corresponding target characteristic value and a corresponding computed characteristic value.

Abstract: Various methods, apparatuses/systems, and media for computing change-agnostic data points are disclosed. A processor trains a machine learning model by using the at least the first set of raw data; computes a set of explanations for all combinations based on output data of the trained machine learning model, the first set of raw data, and sampled raw data computed by applying a sampling algorithm on the raw data; computes a compact representation of the set of explanations corresponding to a pre-configured dimension based on compression quality and generating a set of compressed explanations; computes a unique representation of model explanation with respect to the pre-configured dimension; determines whether the model explanation is robust to changes in data through data perturbation; and computes change-agnostic data points based on determining that the model explanation is robust to changes in data through data perturbation.

Abstract: Various methods, apparatuses/systems, and media for computing strategies for model inferences are disclosed. A processor generates background data from raw data and data sampling strategies associated with a particular security instrument; computes a model for each pair of raw data and machine learning algorithm; computes model explanation for each pair of the background data and the model; normalizes the computed model explanation by utilizing a predefined algorithm; computes a deep dense representation of each explanation based on the normalized explanation of the computed model; clusters the deep dense representation of each explanation; computes a deep dense representation of explanation of each predicted target data version associated with features recovery; compares the deep dense representation of each explanation with the deep dense representation of explanation of each predicted target data version; and computes a strategy of model selection for each target data version as output.

Abstract: There is provided a method and a system for evaluating a relevance score of a given subset of nodes in a knowledge graph (KG) for purpose of link prediction. An ontology used to generate the KG is obtained and clustered to obtain a set of ontology clusters. A set of vectors having been generated by using an embedding model on the KG is obtained and clustered to obtain a set of vector clusters. Training subgraphs are generated based on the set of ontology clusters and the set of vector clusters by removing subsets of nodes from the KG. Respective prediction models are trained on each training subgraph and ranked based on their link predictions. The relevance score of each removed subset of nodes is determined based on the ranked models. A given subset of nodes is provided as a potential explanation based on the relevance score.

Abstract: There is provided a method and system for training an object recognition machine learning model to perform object recognition in data acquired by ultrawide field of view (UW FOV) sensors to thereby obtain a distortion-aware object recognition model. The object recognition model comprises convolution layers each associated with a set of kernels. During training on a UW FOV labelled training dataset, deformable kernels are learned in a manifold space, mapped back to Euclidian space and used to perform convolutions to obtain output feature maps which are used to perform object recognition predictions. Model parameters of the distortion-aware object recognition model may be transferred to other architectures of object recognition models, which may be further compressed for deployment on embedded systems such as electronic devices on board autonomous vehicles.

Abstract: Implementations include providing, by the PKG platform, an initial knowledge graph based on user-specific data associated with a user, and a domain-specific knowledge graph, receiving, by the PKG platform, data representative of at least one answer provided from the user to a respective question, providing, by the PKG platform, an expanded knowledge graph based on the initial knowledge graph, the expanded knowledge graph including one or more nodes and respective edges based on the data, generating, by the PKG platform, a weighted knowledge graph based a groundtruth knowledge graph, and a targeted knowledge graph, the groundtruth knowledge graph including one or more true answers, and the targeted knowledge graph including the at least one answer provided from the user, and generating, by the PKG platform, the hyper-personalized knowledge graph (hpKG) based on the weighted knowledge graph, the hpKG being unique to the user within a domain.

Abstract: A system for predicting the occurrence of an event includes an event detector and a reporting processor. The event detector is configured to: receive data that defines a plurality of social media items; receive a real-time data feed; and predict the occurrence of an event based on a correlation between information in the plurality of social media items and activity associated with the real-time data feed. The reporting processor is configured to determine an event type associated with the event; identify a sentiment of the predicted event based on historical data in the real-time data feed, and generate a recommendation for preventing the occurrence of the event based on at least one of the event type and the sentiment of the predicted event. The recommendation includes a plurality of actions. The reporting processor is coupled to a knowledge graph database that corresponds to an ontology that defines one or more relationships between event types, and response types.

Abstract: Complex computer system architectures are described for utilizing a knowledge data graph comprised of elements, and selecting a discovery element to replace an existing element of a formulation depicted in the knowledge data graph. The substitution process takes advantage of the knowledge data graph structure to improve the computing capabilities of a computing device executing a substitution calculation by translating the knowledge data graph into an embedding space, and determining a discovery element from within the embedding space.

Abstract: There is provided a method and a system for customized image denoising with interpretability. A deep neural network (NN) is trained to denoise an image on a training dataset including pairs of noisy and corresponding clean images acquired from an imaging apparatus, where during the training a structured covariance score (SCS) indicative of a performance of the deep NN in recovering content of corresponding clean images relative to the denoised image is determined based on sparse conditional correlations. A test noisy image is received and denoised by the deep NN. A user feedback score indicative of user satisfaction of the denoising is obtained. A quality parameter is obtained based on the SCS and a quality metric indicative of denoised image quality is obtained from a pretrained NN, and compared with the user feedback score. If the SCS is above the user feedback score, the deep NN is provided for denoising.

Abstract: A system, method and a computer program product may be provided for automatically creating and parameterizing a semantically-enriched diagnosis model for an entity. The system receives a list of data points, from sensors or a database, to be used to create a diagnosis model. The system automatically creates the diagnosis model based on the received list of data points and data stored in a database and parameterizes the diagnosis model. The parameterized diagnosis model reflects rules that determine one or more potential causes of one or more abnormalities of one or more physical conditions in the entity.

Abstract: A method of improving robustness of a deep neural network (DNN), the method including: applying a coverage metric to a trained DNN based on a test set to determine test set adequacy; monitoring a performance of the trained DNN; based on the performance, applying new data to the trained DNN; applying a novelty metric to an output of the trained DNN based on the applied new data to identify a subset of the applied new data in response to determining whether new features are generated; and identifying the subset of the applied new data.

Abstract: Knowledge graph systems are disclosed for enhancing a knowledge graph by generating a new node. The knowledge graph system converts a knowledge graph into an embedding space, and selects a region of interest from within the embedding space. The knowledge graph system further identifies, from the region of interest, one or more gap regions, and calculates a center for each gap region. A node is generated for each gap region, and the information represented by the node is added to the original knowledge graph to generate an updated knowledge graph.

Abstract: A method and a system for generating an augmented scene graph of an image and for training an explainable knowledge based (KB) visual question answering (VQA) machine learning (ML) model are provided. A scene graph encoding spatial and semantic features of objects and relations between objects in the image is obtained. An augmented scene graph is generated by embedding a knowledge graph to enhance the scene graph. An embedded set of questions and associated answers related to the image are obtained. The KB VQA ML model is trained to provide an answer to a given question related to the image based on the augmented scene graph and the embedded set of questions and associated answers. The KB VQA ML model is trained to retrieve a subgraph linking the question and the associated answer as a potential explanation for the answer.

Abstract: A device may determine an association between a second set of parameters and a third set of parameters using a pseudoinversion network and a multiple regression procedure. The device may determine semantic embeddings based on a set of semantic descriptions of the second set of parameters. The device may determine a semantic similarity between parameters of the second set of parameters based on the semantic embeddings. The device may determine a consistency error based on the semantic similarity. The device may generate, using a regression-based learning model technique, a matrix representing an association between the second set of parameters and the third set of parameters based on the association and the consistency error. The device may perform an action based on the matrix.

Abstract: There is provided a method and a system for ranking subgraphs as potential explanations for a labelled edge type class. A first graph representing labelled digital items is obtained, where each labelled digital item is represented as an entity node connected via a labelled edge type to a property value node. The first graph is combined with a second graph representing structured relations in the labelled digital items to obtain a combined graph. Unlabelled digital items are received and matched to respective subgraphs in the combined graph. A machine learning model is used to embed the combined graph to generate graph vectors, and an expressivity score between matched subgraphs and respective labelled edge types based on the generated graph vectors. The matched subgraphs are ranked based on the expressivity score to obtain a ranked set of subgraphs as potential explanations for a respective labelled edge type class.

Abstract: Implementations include providing, by the PKG platform, an initial knowledge graph based on user-specific data associated with a user, and a domain-specific knowledge graph, receiving, by the PKG platform, data representative of at least one answer provided from the user to a respective question, providing, by the PKG platform, an expanded knowledge graph based on the initial knowledge graph, the expanded knowledge graph including one or more nodes and respective edges based on the data, generating, by the PKG platform, a weighted knowledge graph based a groundtruth knowledge graph, and a targeted knowledge graph, the groundtruth knowledge graph including one or more true answers, and the targeted knowledge graph including the at least one answer provided from the user, and generating, by the PKG platform, the hyper-personalized knowledge graph (hpKG) based on the weighted knowledge graph, the hpKG being unique to the user within a domain.

Abstract: Implementations include providing, by the PKG platform, an initial knowledge graph based on user-specific data associated with a user, and a domain-specific knowledge graph, receiving, by the PKG platform, data representative of at least one answer provided from the user to a respective question, providing, by the PKG platform, an expanded knowledge graph based on the initial knowledge graph, the expanded knowledge graph including one or more nodes and respective edges based on the data, generating, by the PKG platform, a weighted knowledge graph based a groundtruth knowledge graph, and a targeted knowledge graph, the groundtruth knowledge graph including one or more true answers, and the targeted knowledge graph including the at least one answer provided from the user, and generating, by the PKG platform, the hyper-personalized knowledge graph (hpKG) based on the weighted knowledge graph, the hpKG being unique to the user within a domain.

Abstract: A method of semantic object detection in an image dataset includes extracting semantic links relevant to the image dataset. Objects are detected in the image dataset and confidence scores are assigned to the detected objects. The semantic object detection compares the detected objects with the semantic links and augments the confidence scores based on the semantic links between the detected objects.

Abstract: Implementations are directed to receiving current data, processing the current data using a predictive model to provide a result, the result corresponding to a sub-model of the predictive model, determining a set of syntactically similar sub-models based on other data, providing at least one semantic model based on the sub-model of the predictive model, one or more syntactically similar sub-models of the set of syntactically similar sub-models, a domain ontology (knowledge graph), and constraints, the at least one semantic model being provided by merging nodes of the sub-model of the predictive model, and a previously determined sub-model of the predictive model using the domain ontology, a label of the domain ontology being used to label a merged node, determining an interpretation based on the at least one semantic model, the interpretation providing at least one reason for the result, and providing the interpretation.