Abstract: An array device and a writing method thereof are provided. A synapse array device includes: a crossbar array, in which a resistive memory element is connected to each intersection of a plurality of row lines and a plurality of column lines; a row select/drive circuit selecting a row line of the crossbar array and applying a pulse signal to the selected row line; a column select/drive circuit selecting a column line of the crossbar array and applying a pulse signal to the selected column line; and a writing part writing to the resistive memory element connected to the selected row line and the selected column line. A first write voltage with controlled pulse width is applied to the selected row line, and a second write voltage with controlled pulse width is applied to the selected column line to perform set writing of the resistive memory element.

Abstract: A first graph that includes a plurality of containers is accessed. The containers each contain one or more rules that each have corresponding computer code. The containers are configured for sequential execution by a rule engine. The computer code corresponding to the one or more rules in each of the containers is electronically scanned. Based on the electronic scan, an interdependency among the rules is determined. Based on the determined interdependency, a second graph is generated. The second graph includes all of the rules of the containers, but not the containers themselves. At least some of the rules are configured for parallel execution by the rule engine.

Abstract: Embodiments are associated with conditional teacher-student model training. A trained teacher model configured to perform a task may be accessed and an untrained student model may be created. A model training platform may provide training data labeled with ground truths to the teacher model to produce teacher posteriors representing the training data. When it is determined that a teacher posterior matches the associated ground truth label, the platform may conditionally use the teacher posterior to train the student model. When it is determined that a teacher posterior does not match the associated ground truth label, the platform may conditionally use the ground truth label to train the student model. The models might be associated with, for example, automatic speech recognition (e.g., in connection with domain adaptation and/or speaker adaptation).

Abstract: There is provided a method, comprising: providing a training dataset including, medical images and corresponding text based reports, and concurrently training a natural language processing (NLP) machine learning (ML) model for generating a NLP category for a target text based report and a visual ML model for generating a visual finding for a target image, by: training the NLP ML model using the text based reports of the training dataset and a ground truth comprising the visual finding generated by the visual ML model in response to an input of the images corresponding to the text based reports of the training dataset, and training the visual ML model using the images of the training dataset and a ground truth comprising the NLP category generated by the NLP ML model in response to an input of the text based reports corresponding to the images of the training dataset.

Abstract: Disclosed is a recognition and training method and apparatus. The apparatus may include a processor configured to input data to a neural network, determine corresponding to a multiclass output a mapping function of a first class and a mapping function of a second class, acquire a result of a loss function including a first probability component that changes correspondingly to a function value of the mapping function of the first class and a second probability component that changes contrastingly to a function value of the mapping function of the second class, determine a gradient of loss corresponding to the input data based on the result of the loss function, update a parameter of the neural network based on the determined gradient of loss for generating a trained neural network based on the updated parameter. The apparatus may input other data to the trained neural network, and indicate a recognition result.

Abstract: Disclosed herein are system, method, and computer program product embodiments for adapting machine learning models for use in additional applications. For example, feature extraction models are readily available for use in applications such as image detection. These feature extraction models can be used to label inputs (such as images) in conjunction with other deep neural network models. However, in adapting the feature extraction models to these uses, it becomes problematic to improve the quality of their results on target data sets, as these feature extraction models are large and resistant to retraining. Approaches disclosed herein include a transfer layer for providing fast retraining of machine learning models.

Abstract: Provided is a system, method, and computer program product for perforated backpropagation. The method includes segmenting a plurality of nodes into at least two sets including a set of first nodes and a set of second nodes, determining an error term for each node of the set of first nodes, the first set of nodes comprising a first and second subset of nodes, backpropagating the error terms for each node throughout the set of first nodes, determining an error term for each node of the first subset of nodes of the set of first nodes based on direct connections between the first subset of nodes and the second subset of nodes independent of error terms of the set of second nodes, determining an error term for each node of the set of second nodes, and updating weights of each node of the plurality of nodes based on the error term.

Abstract: A method for protecting against exposure to content violating a content policy, the method including receiving a number of content items including a first set of content items associated with a content group, determining a measurement associated with an amount of the first set of content items belonging to a specific content category, assigning one or more of the number of content items to be categorized by at least one of the machine learning algorithm or a manual review process, automatically applying the specific content category to one or more other content items of the content group such that the one or more other content items are not reviewed by the manual review process, and transmitting at least one of the number of content items, wherein the content category of each of the number of content items indicates whether the specific content item violates any content policies.

Abstract: Systems and methods are provided for implementing hybrid prediction. Hybrid prediction integrates two deep learning based trajectory prediction approaches: grid-based approaches and graph-based approaches. Hybrid prediction techniques can achieve enhanced performance by combining the grid and graph approaches in a manner that incorporates appropriate inductive biases for different elements of a high-dimensional space. A hybrid prediction framework processor can generate trajectory predictions relating to movement of agents in a surrounding environment based on a prediction model generating using hybrid prediction. Trajectory predictions output from the hybrid prediction framework processor can be used to control an autonomous vehicle. For example, the autonomous vehicle can perform safety-aware and autonomous operations to avoid oncoming objects, based on the trajectory predictions.

Abstract: Systems and methods for processing natural language statements. Based on historical records of data associated with an entity, systems and methods provide models for inferring publication of data content associated with the particular entity. The systems and methods may compare newly observed data content to predicted content associated with an entity for evaluating novelty or impact of the newly observed data content.

Abstract: In a system for storing in memory a tensor that includes at least three modes, elements of the tensor are stored in a mode-based order for improving locality of references when the elements are accessed during an operation on the tensor. To facilitate efficient data reuse in a tensor transform that includes several iterations, on a tensor that includes at least three modes, a system performs a first iteration that includes a first operation on the tensor to obtain a first intermediate result, and the first intermediate result includes a first intermediate-tensor. The first intermediate result is stored in memory, and a second iteration is performed in which a second operation on the first intermediate result accessed from the memory is performed, so as to avoid a third operation, that would be required if the first intermediate result were not accessed from the memory.

Abstract: A processor-implemented neural network operating method, the operating method comprising obtaining a neural network pre-trained in a source domain and a first style feature of the source domain, extracting a second style feature of a target domain from received input data of the target domain, using the neural network, performing domain adaptation of the input data, by performing style matching of the input data based on the first style feature of the source domain and the second style feature of the target domain, and processing the style-matched input data, using the neural network.

Abstract: In one embodiment, a method of machine learning and/or image processing for spectral object classification is described. In another embodiment, a device is described for using spectral object classification. Other embodiments are likewise described.

Abstract: A method and system for training a neural network are described. The method includes providing at least one continuously differentiable model of the neural network. The at least one continuously differentiable model is specific to hardware of the neural network. The method also includes iteratively training the neural network using the at least one continuously differentiable model to provide at least one output for the neural network. Each iteration uses at least one output of a previous iteration and a current continuously differentiable model of the at least one continuously differentiable model.

Abstract: Embodiments relate to an intelligent computer platform to multi-dimensionally optimize device operation. Static hardware device data are acquired and dynamic hardware characteristic data are tracked over one or more temporal segments. A neural model (NM) is trained with corresponding device and network data. The acquired static and dynamic data are input into the NM, and locale processing patterns corresponding to the inputted data are identified. One or more data points and corresponding measurements of the tracked dynamic hardware characteristic data are temporally analyzed. A processing locale corresponding to the temporal analysis is identified and returned as output data, and one or more encoded actions in compliance with the identified processing locale are selectively implemented.

Abstract: In various embodiments, a relevance application quantifies how a trained neural network operates. In operation, the relevance application generates a set of input distributions based on a set of input points associated with the trained neural network. Each input distribution is characterized by a mean and a variance associated with a different neuron included in the trained neural network. The relevance application propagates the set of input distributions through a probabilistic neural network to generate at least a first output distribution. The probabilistic neural network is derived from at least a portion of the trained neural network. Based on the first output distribution, the relevance application computes a contribution of a first input point included in the set of input points to a difference between a first output point associated with a first output of the trained neural network and an estimated mean prediction associated with the first output.

Abstract: Methods, computer systems, and computer-storage medium are provided for providing closed-loop intelligence. A selection of data is received, at a cloud service, from a database comprising data from a plurality of sources in a Fast Healthcare Interoperability Resources (FHIR) format to build a data model. After a feature vector corresponding to the data model is extracted, a selection of an algorithm for a machine learning model to apply to the data model is received. A portion of the selection of data is utilized for training data and test data and the machine learning model is applied to the training data. Once the model is trained, the trained machine learning model can be saved at the cloud service, where it may be accessed by others.

Abstract: Methods, systems, and apparatus, including computer programs encoded on computer storage media, for training a deep neural network. One of the methods includes generating a plurality of feature vectors that each model a different portion of an audio waveform, generating a first posterior probability vector for a first feature vector using a first neural network, determining whether one of the scores in the first posterior probability vector satisfies a first threshold value, generating a second posterior probability vector for each subsequent feature vector using a second neural network, wherein the second neural network is trained to identify the same key words and key phrases and includes more inner layer nodes than the first neural network, and determining whether one of the scores in the second posterior probability vector satisfies a second threshold value.

Abstract: In some implementations there may be provided a system. The system may include a processor and a memory. The memory may include program code which causes operations when executed by the processor. The operations may include analyzing a series of events contained in received data. The series of events may include events that occur during the execution of a data object. The series of events may be analyzed to at least extract, from the series of events, subsequences of events. A machine learning model may determine a classification for the received data. The machine learning model may classify the received data based at least on whether the subsequences of events are malicious. The classification indicative of whether the received data is malicious may be provided. Related methods and articles of manufacture, including computer program products, are also disclosed.

Abstract: In one embodiment, a server instructs one or more networking devices in a local area network (LAN) to form a virtual network overlay in the LAN that redirects traffic associated with a particular node in the LAN to the server. The server receives the redirected traffic associated with the particular node. The server trains a machine learning-based behavioral model for the particular node based on the redirected traffic. The server controls whether a particular redirected traffic flow associated with the node in the LAN is sent to a destination of the traffic flow using the trained behavioral model.

Abstract: A method may include classifying a text by applying a dense convolutional neural network trained to classify the text. The dense convolutional neural network may include one or more dense convolution blocks, each of which including a plurality of convolution layers. Each dense convolution block may be configured to operate on a different quantity of consecutive tokens from the text. Moreover, each of the plurality of convolution layers in a dense convolution block may operate an input to the dense convolution block as well as an output from all preceding convolution layers in the dense convolution block. The text may correspond to an issue associated with a service ticket system. A response for addressing the issue associated with the test may be determined based on the classification of the text. Related systems and articles of manufacture are also provided.

Abstract: A modulation device includes at least one memristive device, and a control block, the modulation device having an equivalent conductance yi(t) produced by the at least one memristive device and the control block being configured to receive a clock signal and perform a first modification of the equivalent conductance yi(t) upon receipt of each clock signal, receive an input voltage pulse and perform a second modification of the equivalent conductance yi(t) upon receipt of each input voltage pulse, the first and second modifications being in opposite directions.

Abstract: A computer-implemented method according to one embodiment includes receiving a training data set to be applied to a model; selecting a subset of the training data set as a sample set; for each of a plurality of predetermined augmentations, applying the predetermined augmentation to the sample set to create an augmented sample set, training the model with the augmented sample set, determining a performance of the trained model, and assigning a weight to the predetermined augmentation for the training data set, based on the determined performance; and selecting one or more of the plurality of predetermined augmentations to be applied to the training data set before the training data set is applied to the model, based on the weight assigned to each of the plurality of predetermined augmentations.

Abstract: Respective labels indicative of compression-related quality degradation for a set of media object tuples which meet a divergence criterion are obtained; each tuple comprises a reference media object and a pair of corresponding compressed media object versions. Pairs of training records for a machine learning model are generated using the labeled media object tuples and multiple perceptual quality algorithms, with each training record comprising respective perceived quality degradation scores generated by each of the multiple algorithms for a given compressed media object of a tuple. A machine learning model is trained, using the record pairs, to predict quality degradation scores for compressed media objects.

Abstract: A vehicle includes: a plurality of sensor devices that determine a driver state; a driver state determining device that receives detection results from a plurality of sensor devices and determines whether the driver state is a dangerous state; and a driving assistance device that performs lane keeping control and speed control of a vehicle and transmits a network connection request to a management server when the driver state determining device has determined the dangerous state.

Abstract: Systems and techniques that facilitate freeze-out as a regularizer in training neural networks are presented. A system can include a memory and a processor that executes computer executable components. The computer executable components can include: an assessment component that identifies units of a neural network, a selection component that selects a subset of units of the neural network, and a freeze-out component that freezes the selected subset of units of the neural network so that weights of output connections from the frozen subset of units will not be updated for a training run.

Abstract: In some forms containing keywords and content, there may be nested levels of keywords, also referred to as a hierarchy. Content in the forms may be associated with one or more keywords in one or more of the nested levels, or in the hierarchy. Identifying keywords in adjacent cells in a table (with a nested keyword being either to the right of or below another keyword) enables distinguishing between keywords and content in filled forms, and enables correct association of content with respective keywords.

Abstract: A system, a method and a computer program product are provided for evaluating quality of data, such as sensor data and map data, using a machine learning model. The system may include at least one memory configured to store computer executable instructions and at least one processor configured to execute the computer executable instructions to obtain first sensor features of the first sensor data associated with a road object in a first geographic region, first map features of the first map data associated with the road object and ground truth data associated with the road object. The processor may be configured to generate the machine learning model by configuring the ground truth data and calculating first information scores for each of the first sensor features and the first map features by recursively splitting each of the first sensor features and the first map features.

Abstract: An information processing device which performs semi-supervised learning is provided with: a dictionary input circuit for acquiring a dictionary, a parameter group used by an identification device; a boundary determination circuit which obtains an identification boundary for the dictionary on the basis of the dictionary, supervised data, and labelled unsupervised data; a labelling circuit which labels the unsupervised data in accordance with the identification boundary; a loss calculation circuit which calculates the sum total of supervised-data loss calculated in accordance with the labels assigned in advance and the labels based on the identification boundary, and unsupervised-data loss calculated such that further from the identification boundary the smaller the loss; a dictionary update circuit which updates the dictionary such that the sum-total loss is reduced; and a dictionary output circuit which outputs the updated dictionary.

Abstract: In some implementations, a system is capable of obtaining and processing both actively monitored and passively monitored data in parallel in order to improve the accuracy and the specificity by which pathological risks are identified for a user. Data indicating measured levels of one or more metabolic biomarkers and activity data associated with a user is obtained. A biological state for the user is determined based on the measured levels of the one or more metabolic biomarkers. One or more user inputs indicated within the activity data, and scores reflecting respective likelihoods that a particular user input indicates a change to one or more aspects of the biological state for the user for each of the one or more user inputs is determined. Data corresponding to the biological state for the user is then adjusted. A communication that is generated based on the adjusted data is then provided for output.

Abstract: A processor-implemented method includes determining a first quantization value by performing log quantization on a parameter from one of input activation values and weight values in a layer of a neural network, comparing a threshold value with an error between a first dequantization value obtained by dequantization of the first quantization value and the parameter, determining a second quantization value by performing log quantization on the error in response to the error being greater than the threshold value as a result of the comparing; and quantizing the parameter to a value in which the first quantization value and the second quantization value are grouped.

Abstract: Disclosed is a recognition and training method and apparatus. The apparatus may include a processor configured to input data to a neural network, determine corresponding to a multiclass output a mapping function of a first class and a mapping function of a second class, acquire a result of a loss function including a first probability component that changes correspondingly to a function value of the mapping function of the first class and a second probability component that changes contrastingly to a function value of the mapping function of the second class, determine a gradient of loss corresponding to the input data based on the result of the loss function, update a parameter of the neural network based on the determined gradient of loss for generating a trained neural network based on the updated parameter. The apparatus may input other data to the trained neural network, and indicate a recognition result.

Abstract: A live stream, that includes a video stream and an audio stream, of a presenter is monitored. The live stream is attended by an audience that includes one or more audience members. One or more stream content features of the live stream at a first window of time is transmitted to a multimodal machine learning model. One or more audience content features of the audience at the first window of time is transferred to the multimodal model. One or more feature results, based on the stream content features and based on the audience content features, of the first window of time is obtained from the multimodal model. The feature results are sent to an auditory machine learning model. A first audio signal from the auditory machine learning model is received. An augmented stream of the first window of time is generated based on the first audio signal.

Abstract: Methods, systems, and apparatus, including computer programs encoded on computer storage medium, for training a neural network, wherein the neural network is configured to receive an input data item and to process the input data item to generate a respective score for each label in a predetermined set of multiple labels. The method includes actions of obtaining a set of training data that includes a plurality of training items, wherein each training item is associated with a respective label from the predetermined set of multiple labels; and modifying the training data to generate regularizing training data, comprising: for each training item, determining whether to modify the label associated with the training item, and changing the label associated with the training item to a different label from the predetermined set of labels, and training the neural network on the regularizing data.

Abstract: Machine learning classification models which are robust against label noise are provided. Noise may be modelled explicitly by modelling “label flips”, where incorrect binary labels are “flipped” relative to their ground truth value. Distributions of label flips may be modelled as prior and posterior distributions in a flexible architecture for machine learning systems. An arbitrary classification model may be provided within the system. The classification model is made more robust to label noise by operation of the prior and posterior distributions. Particular prior and approximating posterior distributions are disclosed.

Abstract: Disclosed is an artificial intelligence (AI) server. The AI server includes a communication unit configured to communicate with an AI device; and an AI unit configured to receive feature data from the AI device, wherein the received feature data is generated by the AI device by obtaining sensing data and compressing the sensing data while preserving a feature of the sensing data; and input the received feature data to a deep learning model to obtain second sensing data for use in a recognition model related to an AI function of the AI device.

Abstract: A system for analyzing machine learning-derived misappropriation types with an array of shadow models is provided. The system comprises: a controller configured for analyzing an output of a machine learning model, the controller being further configured to: input interaction data into a machine learning model, wherein the interaction data is analyzed using the machine learning model to determine a misappropriation type output associated with the interaction data; identify data features in the interaction data associated with the misappropriation type output; construct an array of shadow models based on the data features, wherein each individual model in the array of shadow models is configured to extract logical constructs from a portion of the data features; and consolidate the logical constructs output by the array of shadow models, wherein consolidating the logical constructs determines a final explanation output for the misappropriation type output determined by the machine learning model.

Abstract: Systems and methods for aggregating data. The system is configured to receive metadata from an interactive graphical user interface (GUI) of a user device, aggregate field values from the data stored on one or more databases based on the received metadata and generate filter instructions based on the received metadata. The system is further configured to transmit the aggregated field values and the filter instructions to the user device, receive a user-customized filter set and subscription request for a synthetic symbol associated with the user-customized filter set from the user device, and create the synthetic symbol responsive to the subscription request. Moreover, the system aggregates one or more data values from the data stored on the databases associated with the created synthetic symbol and generates instructions to display the data values on the interactive GUI in accordance with the user-customized filter set associated with the created synthetic symbol.

Abstract: An example system includes a processor to evaluate a trained first classifier on a test set of labeled data to generate error rates for a number of labels. The processor is to process a set of unlabeled data via the trained first classifier to generate annotated data including labels and associated error rates. The processor is to train a second classifier using the annotated data and the associated error rates.

Abstract: Systems, architectures, and approaches for use with neural networks. An execution block and a system architecture using a novel execution block are disclosed along with how such an execution block can be used. The execution block uses a fully connected stack of layers and parameters of this fully connected stack of layers are shared. The fully connected nature of the block and on-the-fly generated parameters allow for bypassing specialized training data sets. The system may be trained using non-task-specific training data sets and this allows the system to transfer what is learned to execute a different task. Thus, instead of having to obtain a specialized training data set for a specific task, a more generic training data set can be used to train and prepare the system for that specific task. Results have shown that performance is as good as than the state of the art in providing solutions.

Abstract: To improve the reliability of mutual diagnosis in a cancer determination by machine learning, m/z values of ions originating from tumor markers or similar substances used in other related tests are stored in a particular m/z-value database. A spectrum information filtering section deletes signal intensities at the m/z values stored in the particular m/z-value database from a large number of mass spectra classified by the presence or absence of cancer. Using the data which remain after the deletion as training data, a training processor obtains training-result information and stores it in a training result database. A judgment processor similarly deletes signal intensities at the predetermined m/z values from mass spectrum data obtained for a target sample to be judged. Then, based on the training-result information stored in the training-result database, the judgment processor determines whether the target sample should be classified into a cancerous group or non-cancerous group.

Abstract: Implementations can identify a given assistant device from among a plurality of assistant devices in an ecosystem, obtain device-specific signal(s) that are generated by the given assistant device, process the device-specific signal(s) to generate candidate semantic label(s) for the given assistant device, select a given semantic label for the given semantic device from among the candidate semantic label(s), and assigning, in a device topology representation of the ecosystem, the given semantic label to the given assistant device. Implementations can optionally receive a spoken utterance that includes a query or command at the assistant device(s), determine a semantic property of the query or command matches the given semantic label to the given assistant device, and cause the given assistant device to satisfy the query or command.

Abstract: A computer system trains a machine learning model. A vector representation is generated for each document in a collection of documents. The documents are clustered based on the vector representations of the documents to produce a plurality of clusters. A training set is produced by selecting one or more documents from each cluster, wherein the selected documents represent a sample of the collection of documents to train the machine learning model. The machine learning model is trained by applying the training set to the machine learning model. Embodiments of the present invention further include a method and program product for training a machine learning model in substantially the same manner described above.

Abstract: Disclosed is a method and apparatus for inspecting defects in a washer based on deep learning. According to an embodiment of the present disclosure, a method for inspecting defects in a washer based on deep learning gathers learning data while the washer operates and trains a first ANN model for diagnosing the condition of the washer and a second ANN model for securing the reliability of the result of inspection of the condition of the washer. Thereafter, the washer may make a diagnosis of whether the washer is defective based on the two pre-trained ANN models and are thereby able to continuously monitor whether the washer has an abnormal condition. According to an embodiment, the artificial intelligence (AI) module may be related to unmanned aerial vehicles (UAVs), robots, augmented reality (AR) devices, virtual reality (VR) devices, and 5G service-related devices.

Abstract: Methods, systems, and apparatus for a method that predicts an individual survival survival time of a patient. The method includes obtaining clinical data associated with health factors of the patient. The method includes obtaining liquid biopsy data associated with one or more attributes of diseased cells within the patient. The method includes predicting or determining a survival time of the patient using a deep learning model based on the clinical data and the liquid biopsy data. The method includes providing or outputting the survival time.

Abstract: Systems, methods, and apparatuses for generating and using machine learning models using genetic data. A set of input features for training the machine learning model can be identified and used to train the model based on training samples, e.g., for which one or more labels are known. As examples, the input features can include aligned variables (e.g., derived from sequences aligned to a population level or individual references) and/or non-aligned variables (e.g., sequence content). The features can be classified into different groups based on the underlying genetic data or intermediate values resulting from a processing of the underlying genetic data. Features can be selected from a feature space for creating a feature vector for training a model. The selection and creation of feature vectors can be performed iteratively to train many models as part of a search for optimal features and an optimal model.

Abstract: The computer-implemented or hardware-implemented method of entity identification, comprising: a) providing a network of nodes with input from a plurality of sensors; b) generating, by each node of the network, an activity level, based on the input from the plurality of sensors; c) comparing the activity level of each node to a threshold level; d) based on the comparing, for each node, setting the activity level to a preset value or keeping the generated activity level; e) calculating a total activity level as the sum of all activity levels of the nodes of the network; f) iterating a)-e) until a local minimum of the total activity level has been reached; and g) when the local minimum of the total activity level has been reached, utilizing a distribution of activity levels at the local minimum to identify a measurable characteristic of the entity. The disclosure further relates to a computer program product and an apparatus for entity identification.

Abstract: A set of distance measurable encrypted feature vectors can be derived from any biometric data and/or physical or logical user behavioral data, and then using an associated deep neural network (“DNN”) on the output (i.e., biometric feature vector and/or behavioral feature vectors, etc.) an authentication system can determine matches or execute searches on encrypted data. Behavioral or biometric encrypted feature vectors can be stored and/or used in conjunction with respective classifications, or in subsequent comparisons without fear of compromising the original data. In various embodiments, the original behavioral and/or biometric data is discarded responsive to generating the encrypted vectors. In another embodiment, distance measurable or homomorphic encryption enables computations and comparisons on cypher-text without decryption of the encrypted feature vectors. Security of such privacy enabled embeddings can be increased by implementing an assurance factor (e.g.

Abstract: A system may automatically create training datasets for training a segmentation model to recognize features such as lanes on a road. The system may receive sensor data representative of a portion of an environment and map data from a map data store including existing map data for the portion of the environment that includes features present in that portion of the environment. The system may project or overlay the features onto the sensor data to create training datasets for training the segmentation model, which may be a neural network. The training datasets may be communicated to the segmentation model to train the segmentation model to segment data associated with similar features present in different sensor data. The trained segmentation model may be used to update the map data store, and may be used to segment sensor data obtained from other portions of the environment, such as portions not previously mapped.

Abstract: Fusion of trained artificial intelligence (AI) neural networks to produce more accurate classifications is disclosed. Concatenation from each network being merged may be performed. The new set of features, which includes the concatenated layers, is then fed through a new classifier to form a single final classifier that uses the best parts of each input classifier.