Abstract: An off-policy reinforcement learning actor-critic neural network system configured to select actions from a continuous action space to be performed by an agent interacting with an environment to perform a task. An observation defines environment state data and reward data. The system has an actor neural network which learns a policy function mapping the state data to action data. A critic neural network learns an action-value (Q) function. A replay buffer stores tuples of the state data, the action data, the reward data and new state data. The replay buffer also includes demonstration transition data comprising a set of the tuples from a demonstration of the task within the environment. The neural network system is configured to train the actor neural network and the critic neural network off-policy using stored tuples from the replay buffer comprising tuples both from operation of the system and from the demonstration transition data.

Abstract: Disclosed herein are methods, systems, and computer-readable media for training a machine learning model to label unlabeled data and/or perform automated actions. In an embodiment, a method comprises receiving unlabeled digital video data, generating pseudo-labels for the unlabeled digital video data, the generating comprising receiving labeled digital video data, training an inverse dynamics model (IDM) using the labeled digital video data, and generating at least one pseudo-label for the unlabeled digital video data, wherein the at least one pseudo-label is based on a prediction, generated by the IDM, of one or more actions that mimic at least one timestep of the unlabeled digital video data. In some embodiments, the method further comprises adding the at least one pseudo-label to the unlabeled digital video data and further training the IDM or a machine learning model using the pseudo-labeled digital video data.

Abstract: In one embodiment, a method includes obtaining candidate data generated by a vehicle. The candidate data comprises a subset of sensor data identified based on a set of neural network models executing on the vehicle. The method also includes determining whether the candidate data can be associated with one or more categories of a set of categories for training data based on a set of categorization models. The method further includes associating the candidate data with the first category in response to determining that the candidate data can be associated with at a first category of the set of categories. The method further includes determining whether the candidate data can be associated with a second category. The set of categories lacks the second category. The method further includes including the second category in the set of categories in response to determining that the candidate data can be associated with the second category.

Abstract: A system and method is disclosed for automated cross-node communication in a distributed directed acyclic graph. The method can include identifying a directed acyclic graph (“DAG”) overlaying a plurality of nodes and identifying the nodes underlying the DAG. A subordinate DAG can be generated in an entry vertex of the DAG. The subordinate DAG can include a vertex for each of the nodes underlying the DAG. Data and metadata can be received at the entry vertex. The data can be delivered to a next vertex in the DAG, and the metadata can be communicated to nodes underlying the DAG via the subordinate DAG.

Abstract: A memory controller includes an interface and a processor. The interface communicates with memory cells organized in multiple Word Lines (WLs). The processor is configured to read a Code Word (CW) of an Error Correction Code (ECC) from a group of multiple memory cells belonging to a target WL, to calculate for a given memory cell (i) a first soft metric, depending on a first threshold voltage of a first neighbor memory cell in a first WL neighboring the target WL, and (ii) a second soft metric, depending on a second threshold voltage of a second neighbor memory cell in a second WL neighboring the target WL, to calculate a combined soft metric based on both the first and second soft metrics and assign the combined soft metric to the given memory cell, and to decode the CW based on the combined soft metric, to produce a decoded CW.

Abstract: One embodiment includes a method for generating representations of inputs with missing values. The method includes steps for receiving an input includes a set of one or more values for several features, wherein the set of values for at least one of the several features includes values for each of several points in time, and for identifying a missingness pattern of the input, wherein the missingness pattern for the at least one feature indicates whether the set of values is missing a value for each of the several points in time. The method further includes steps for determining a set of one or more transformation weights based on the missingness pattern, and transforming the input based on the determined transformation weights.

Abstract: Embodiments are generally directed to an adaptive deformable kernel prediction network for image de-noising. An embodiment of a method for de-noising an image by a convolutional neural network implemented on a compute engine, the image including a plurality of pixels, the method comprising: for each of the plurality of pixels of the image, generating a convolutional kernel having a plurality of kernel values for the pixel; generating a plurality of offsets for the pixel respectively corresponding to the plurality of kernel values, each of the plurality of offsets to indicate a deviation from a pixel position of the pixel; determining a plurality of deviated pixel positions based on the pixel position of the pixel and the plurality of offsets; and filtering the pixel with the convolutional kernel and pixel values of the plurality of deviated pixel positions to obtain a de-noised pixel.

Abstract: This disclosure involves generating graph data structures that model inter-feature dependencies for use with machine-learning models to predict end-user behavior. For example, a processing device receives an input dataset and a request to modify a first input feature of the input dataset. The processing device uses a graph data structure that models the inter-feature dependencies to modify the input dataset by propagating the modification of the first input feature to a second input feature dependent on the first input feature. The modification to the second input feature is a function of at least (a) the value of the first input feature and (b) a weight assigned to an edge linking the first input feature to the second input feature within the directed graph. The processing device then applies a trained machine-learning model to the modified input dataset to generate a prediction of an outcome.

Abstract: In order to enable stabilizing control of communication in a communication network, a system according to an aspect of the present disclosure includes: an obtaining means for obtaining work-related information related to human work in network operation; and a training means for training a machine learning based controller for controlling communication in a communication network, based on the work-related information.

Abstract: The present disclosure provides an image classification method, apparatus, and device, and a readable storage medium. The image classification method includes: processing an image to be processed, by using a first convolutional network, to obtain a first feature map; processing the first feature map, by using a residual network, to obtain a second feature map, wherein the residual network includes a depth separable convolutional layer; and processing the second feature map, by using a second convolutional network, to determine a category label of the image to be processed.

Abstract: A system and method are disclosed for providing a private multi-modal artificial intelligence platform. The method includes splitting a neural network into a first client-side network, a second client-side network and a server-side network and sending the first client-side network to a first client. The first client-side network processes first data from the first client, the first data having a first type. The method includes sending the second client-side network to a second client. The second client-side network processes second data from the second client, the second data having a second type. The first type and the second type have a common association. Forward and back propagation occurs between the client side networks and disparate data types on the different client side networks and the server-side network to train the neural network.

Abstract: An endoscopic image processing method is provided. The method can include acquiring a current endoscopic image of a to-be-examined user, and predicting the current endoscopic image by using a deep convolutional network based on a training parameter. The training parameter can be determined according to at least one first endoscopic image and at least one second endoscopic image transformed from the at least one first endoscopic image, where the at least one endoscopic image corresponds to a human body part. The method can further include determining an organ category corresponding to the current endoscopic image. The method can make a prediction process more intelligent and more robust, thereby improving resource utilization of a processing apparatus.

Abstract: A data structuring system that provides a user interface to enable data wrangling and modeling, and methods for making and using the same.

Abstract: Embodiments for providing an enhanced codebase in a computing environment by a processor. One or more container specification files may be automatically generated for a codebase based on one or more extracted attribute names and values.

Abstract: Systems as described herein may label data to preserve privacy. An annotation server may receive a document comprising a collection of text representing a plurality of confidential data from a first computing device. The annotation server may convert the document to a plurality of text embeddings. The annotation server may input the text embeddings into a machine learning model to generate a plurality of synthetic images, and receive a label for each of the plurality of synthetic images from a third-party labeler. Accordingly, the annotation server may send the confidential data and the corresponding labels to a second computing device.

Abstract: Computer systems and computer-implemented methods modify a machine learning network, such as a deep neural network, to introduce judgment to the network. A “combining” node is added to the network, to thereby generate a modified network, where activation of the combining node is based, at least in part, on output from a subject node of the network. The computer system then trains the modified network by, for each training data item in a set of training data, performing forward and back propagation computations through the modified network, where the backward propagation computation through the modified network comprises computing estimated partial derivatives of an error function of an objective for the network, except that the combining node selectively blocks back-propagation of estimated partial derivatives to the subject node, even though activation of the combining node is based on the activation of the subject node.

Abstract: Embodiments for ensemble training in a distributed marketplace in a computing environment. One or more ensemble machine learning models may be provided from a plurality of machine learning models competing within the distributed marketplace that achieve a performance on ensemble training data equal to or greater than a selected performance threshold, wherein the distributed marketplace is a blockchain.

Abstract: The display device includes a plurality of pixels, a data driver, and a timing controller. The plurality of pixels are connected to a plurality of scan lines and a plurality of data lines. The data driver supplies a data voltage to the plurality of data lines in a light-emitting mode and supplies a neural network input voltage to the plurality of data lines in an artificial neural network mode. The timing controller is in an artificial neural network mode, and supplies a weight value control signal for performing a deep learning operation by using at least one of the plurality of pixels to the data driver. The weight value control signal is generated based on a predetermined weight value.

Abstract: A method for programming an activation function is provided. The method includes generating segment data for segmenting the activation function; segmenting the activation function into a plurality of segments using the segment data; and approximating at least one segment of the plurality of segments as a programmable segment. An apparatus for performing the method may include a programmable activation function generator configured to generate segment data for segmenting an activation function; segment the activation function into a plurality of segments using the generated segment data; and approximate at least one segment of the plurality of segments as a programmable segment. By using segment data, various non-linear activation functions, particularly newly proposed or known activation functions with some modifications, can be programmed to be processable in hardware.

Abstract: A diagnostic system for model governance is presented. The diagnostic system includes an auto-encoder to monitor model suitability for both supervised and unsupervised models. When applied to unsupervised models, the diagnostic system can provide a reliable indication on model degradation and recommendation on model rebuild. When applied to supervised models, the diagnostic system can determine the most appropriate model for the client based on a reconstruction error of a trained auto-encoder for each associated model. An auto-encoder can determine outliers among subpopulations of consumers, as well as support model go-live inspections.

Abstract: A processing apparatus includes a driver configured to drive a controlled object, and a controller configured to control the driver by generating a command value to the driver based on a control error. The controller includes a first compensator configured to generate a first command value based on the control error, a second compensator configured to generate a second command value based on the control error, and an adder configured to obtain the command value by adding the first command value and the second command value. The second compensator includes a neural network for which a parameter value is decided by learning, and input parameters input to the neural network include at least one of a driving condition of the driver and an environment condition in a periphery of the controlled object in addition to the control error.

Abstract: Disclosed herein is technology for performing predictive modeling to identify inputs for a manufacturing process. An example method may include receiving expected output data for a manufacturing process, wherein the expected output data defines an attribute of an output of the manufacturing process; accessing a plurality of machine learning models that model the manufacturing process; determining, using a first machine learning model, input data for the manufacturing process based on the expected output data for the manufacturing process, wherein the input data comprises a value for a first input and a value for a second input; combining the input data determined using the first machine learning model with input data determined using the second machine learning model to produce a set of inputs for the manufacturing process, wherein the set of inputs comprises candidate values for the first input and candidate values for the second input.

Abstract: [Object] To provide a lightning threat information-providing apparatus, a lightning threat information-providing method, and a program that are capable of providing a user with accurate information regarding a lightning threat.

Abstract: Systems and methods for natural language processing are described. One or more embodiments of the disclosure provide an entity matching apparatus trained using machine learning techniques to determine whether a query name corresponds to a candidate name based on a similarity score. In some examples, the query name and the candidate name are encoded using a character encoder to produce a regularized input sequence and a regularized candidate sequence, respectively. The regularized input sequence and the regularized candidate sequence are formed from a regularized character set having fewer characters than a natural language character set.

Abstract: A segmentation system utilizes a supervised learning method and a clustering analysis to identify clusters, thereby segmenting a population into groups, where the clusters are associated with various conversion potentials that indicate the probability of an event. The segmentation system employs the supervised learning method to train a model on training data comprising historical conversion data and features associated with members of the group. A subset of features is selected from a ranked order that is determined using weights generated by the supervised learning. A clustering analysis is performed for a population with respect to the subset to generate clusters. A superior cluster is identified based on it having a conversion potential greater than a conversion potential of another cluster. In a marketing context, the system can be employed to identify a superior cluster of users that have a higher conversion potential in response to an advertisement campaign.

Abstract: Certain aspects of the present disclosure are directed to methods and apparatus for configuring a multiply-accumulate (MAC) block in an artificial neural network. A method generally includes receiving, at a neural processing unit comprising one or more logic elements, at least one input associated with a use-case of the neural processing unit; obtaining a set of weights associated with the at least one input; selecting a precision for the set of weights; modifying the set of weights based on the selected precision; and generating an output based, at least in part, on the at least one input, the modified set of weights, and an activation function.

Abstract: Systems, methods and computer program products for improving security of artificial intelligence systems. The system comprising processors for monitoring one or more transactions received by a machine learning decision model to determine a first score associated with a first transaction. The first transaction may be identified as likely adversarial, in response to the first score being lower than a certain score threshold and the first transaction having a low occurrence likelihood. A second score may be generated in association with the first transaction based on one or more adversarial latent features associated with the first transaction. At least one adversarial latent feature may be detected as being exploited by the first transaction, in response to determining that the second score falls above the certain score threshold. Accordingly, an abnormal volume of activations of adversarial latent features spanning across a plurality of transactions scored may be detected and blocked.

Abstract: A computer-implemented method includes receiving, by a computing device, input activations and determining, by a controller of the computing device, whether each of the input activations has either a zero value or a non-zero value. The method further includes storing, in a memory bank of the computing device, at least one of the input activations. Storing the at least one input activation includes generating an index comprising one or more memory address locations that have input activation values that are non-zero values. The method still further includes providing, by the controller and from the memory bank, at least one input activation onto a data bus that is accessible by one or more units of a computational array. The activations are provided, at least in part, from a memory address location associated with the index.

Abstract: A method for modularizing high dimensional neural networks into neural networks of lower input dimensions. The method is suited to generating full-DOF robot grasping actions based on images of parts to be picked. In one example, a first network encodes grasp positional dimensions and a second network encodes rotational dimensions. The first network is trained to predict a position at which a grasp quality is maximized for any value of the grasp rotations. The second network is trained to identify the maximum grasp quality while searching only at the position from the first network. Thus, the two networks collectively identify an optimal grasp, while each network's searching space is reduced. Many grasp positions and rotations can be evaluated in a search quantity of the sum of the evaluated positions and rotations, rather than the product. Dimensions may be separated in any suitable fashion, including three neural networks in some applications.

Abstract: Items within an index may be converted from classic geometry and embedded into a hyperbolic space. The hyperboloids within the hyperbolic space provide higher precision classifications of items within the index relative to their hierarchical structure. A received search query may also be converted to hyperbolic space and mapped as a query hyperboloid against an answer space that includes hyperboloids for items within the index. Distances or overlaps between the hyperboloids may be determined in order to generate a set of search results.

Abstract: Hardware for implementing a Deep Neural Network (DNN) having a convolution layer, the hardware comprising a plurality of convolution engines each configured to perform convolution operations by applying filters to data windows, each filter comprising a set of weights for combination with respective data values of a data window; and one or more weight buffers accessible to each of the plurality of convolution engines over an interconnect, each weight buffer being configured to provide weights of one or more filters to any of the plurality of convolution engines; wherein each of the convolution engines comprises control logic configured to request weights of a filter from the weight buffers using an identifier of that filter.

Abstract: The present technology is effective to cause at least one processor to instruct an autonomous vehicle to navigate a specific course and to record diagnostic measurements while navigating the specific course, receive the diagnostic measurements from the autonomous vehicle, and analyze the diagnostic measurements from the autonomous vehicle in a context provided by a collection of diagnostic measurement data collected from a fleet of similar autonomous vehicles navigating the specific course.

Abstract: Methods, systems, apparatus, including computer programs encoded on computer storage media for generating psychiatric treatment recommendations. In one aspect, the method includes actions of receiving, by a server and from a first user device, one or more data structures that collectively include fields structuring data that represents (i) current patient symptoms and (ii) current patient diagnoses, generating, by the server, rendering data structure that includes fields structuring data that represents rendering data that, when rendered on a display device, presents a patient dashboard based on the data that is received from the mobile device, and providing, by the server, the rendering data structure to a second user device that is different than the first user device, wherein the second user device is configured to render the rendering data structured by the fields of the rendering data structure to output the patient dashboard on the display of the second user device.

Abstract: Quantization-aware neural architecture search (“QNAS”) can be utilized to learn optimal hyperparameters for configuring an artificial neural network (“ANN”) that quantizes activation values and/or weights. The hyperparameters can include model topology parameters, quantization parameters, and hardware architecture parameters. Model topology parameters specify the structure and connectivity of an ANN. Quantization parameters can define a quantization configuration for an ANN such as, for example, a bit width for a mantissa for storing activation values or weights generated by the layers of an ANN. The activation values and weights can be represented using a quantized-precision floating-point format, such as a block floating-point format (“BFP”) having a mantissa that has fewer bits than a mantissa in a normal-precision floating-point representation and a shared exponent.

Abstract: A computer-implemented method for providing a simulation model of an electric rotating machine is disclosed. The simulation model is defined by parameter values. Input data is obtained. The input data is collectable using the electric rotating machine when the electric rotating machine is not connected to an operating voltage and being characteristic of the electric rotating machine. The parameter values are determined from the input data using a trained function and the parameter values determined are provided.

Abstract: Disclosed are a defense method and a model of deep learning model aiming at adversarial attacks in the technical field of image recognition, which makes full use of the internal relationship between the adversarial samples and the initial samples, and transforms the adversarial samples into common samples by constructing a filter layer in front of the input layer of the deep learning model; the parameters of the filter layer are trained by using the adversarial attack samples, so as to improve the ability of the model to resist adversarial attack; then the trained filter layer is combined with the learning model after the adversarial training, and a deep learning model with strong robustness and high classification accuracy is obtained, which ensures that the recognition ability of the initial sample is not reduced while resisting the adversarial attacks.

Abstract: A neural processor. In some embodiments, the processor includes a first tile, a second tile, a memory, and a bus. The bus may be connected to the memory, the first tile, and the second tile. The first tile may include: a first weight register, a second weight register, an activations buffer, a first multiplier, and a second multiplier. The activations buffer may be configured to include: a first queue connected to the first multiplier and a second queue connected to the second multiplier. The first queue may include a first register and a second register adjacent to the first register, the first register being an output register of the first queue. The first tile may be configured: in a first state: to multiply, in the first multiplier, a first weight by an activation from the output register of the first queue, and in a second state: to multiply, in the first multiplier, the first weight by an activation from the second register of the first queue.

Abstract: An object identification method includes: generating a tracking sample and an adversarial sample; training a teacher model according to the tracking sample; and initializing a student model according to the teacher model. The student model adjusts a plurality of parameters according to the teacher model and the adversarial sample, in response to the vector difference between the output result of the student model and the output result of the teacher model being lower than the learning threshold, the student model is deemed to have completed training, and the student model is extracted as an object identification model.

Abstract: An artificial intelligence (AI) platform to support optimization of container builds and virtual machine mounts in a distributed computing environment. A provisioning file is subject to natural language processing (NLP) and a corresponding vector representation of the file is created and subject to evaluation by a set of artificial neural networks (ANN). A first ANN assesses the representation of the file with respect to compliance and operability, and the second ANN selectively assesses the representation of the file with respect to provisioning efficiency. The provisioning file is selectively process based on the provisioning efficiency, with the processing directed at provisioning a container build or mounting a VM.

Abstract: Embodiments of a method and system for improving care determination for care providers in relation to a condition of a user associated with a mobile device can include: collecting a log of use dataset associated with user digital communication behavior at the mobile device; collecting a mobility supplementary dataset corresponding to a mobility-related sensor of the mobile device; determining a medical status analysis for a condition of the user based on at least one of the log of use dataset and the mobility supplementary dataset, the medical status analysis including at least one of a diagnosis and a therapeutic intervention associated with the condition; and promoting the at least one of the diagnosis and the therapeutic intervention to a care provider.

Abstract: An apparatus for optimizing a computational network is configure to receive an input at a first processing component. The first processing component may include at least a first programmable processing component and a second programmable processing component. The first programmable processing component is configured to compute a first nonlinear function and the second programmable processing component is configured to compute a second nonlinear function which is different than the second nonlinear function. The computational network which may be a recurrent neural network such as a long short-term memory may be operated to generate an inference based at least in part on outputs of the first programmable processing component and the second programmable processing component.

Abstract: Systems and methods for distributed training of deep learning models are disclosed. An example local device to train deep learning models includes a reference generator to label input data received at the local device to generate training data, a trainer to train a local deep learning model and to transmit the local deep learning model to a server that is to receive a plurality of local deep learning models from a plurality of local devices, the server to determine a set of weights for a global deep learning model, and an updater to update the local deep learning model based on the set of weights received from the server.

Abstract: Systems and methods relating to machine learning. An edge device runs a new data point on a first neural network and determines activations on the layers within that neural network. The first neural network is a fully trained network based on a second neural network on a server. The activation data for the various layers in the first neural network are, starting with the output layer, sequentially transmitted to the server. The server continuously receives this activation data and continuously compares it with previously encountered activation data for the second neural network. If the received activation data is within an expected range, then the edge device is instructed to stop sending activation data. Otherwise, the server continues to receive the activation data for the other layers until the new data point is received by the server or the activation data is within the expected range of previously encountered activation data.

Abstract: The present disclosure relates to systems, methods, and non-transitory computer readable media for efficiently, quickly, and flexibly generating and providing pixel-wise classification predictions utilizing early exit heads of a multi-exit pixel-level prediction neural network. For example, the disclosed systems utilize a multi-exit pixel-level prediction neural network to generate classification predictions for a digital image on the pixel level. The multi-exit pixel-level prediction neural network includes a specialized architecture with early exit heads having unique encoder-decoder architectures for generating pixel-wise classification predictions at different early exit stages. In some embodiments, the disclosed systems implement a spatial confidence-adaptive scheme to mask certain predicted pixels to prevent further processing of the masked pixels and thereby reduce computation.

Abstract: A system comprising an electronic device that includes a processor is described. During operation, the processor acquires a full version of a neural network, the neural network including internal elements for processing instances of input image data having a set of color channels. The processor then generates, from the neural network, a set of sub-networks, each sub-network being a separate copy of the neural network with the internal elements for processing at least one of the color channels in instances of input image data removed, so that each sub-network is configured for processing a different set of one or more color channels in instances of input image data. The processor next provides the sub-networks for processing instances of input image data—and may itself use the sub-networks for processing instances of input image data.

Abstract: Methods and systems, including computer programs encoded on a computer storage medium. In one aspect, a method includes the actions of receiving a request to perform convolutional computations for a neural network on a hardware circuit having a matrix computation unit, the request specifying the convolutional computation to be performed on a feature tensor and a filter and padding applied to the feature tensor prior to performing the convolutional computation; and generating instructions that when executed by the hardware circuit cause the hardware circuit to perform operations comprising: transferring feature tensor data from a main memory of the hardware circuit to a scratchpad memory of the hardware circuit; and repeatedly performing the following operations: identifying a current subset of the feature tensor; and determining whether a memory view into the scratchpad memory for the current subset is consistent with a memory view of the current subset in the main memory.

Abstract: A system having multiple devices that can host different versions of an artificial neural network (ANN). In the system, changes to local versions of the ANN can be combined with a master version of the ANN. In the system, a first device can include memory that can store the master version, a second device can include memory that can store a local version of the ANN, and there can be many devices that store local versions of the ANN. The second device (or any other device of the system hosting a local version) can include a processor that can train the local version, and a transceiver that can transmit changes to the local version generated from the training. The first device can include a transceiver that can receive the changes to a local version, and a processing device that can combine the received changes with the master version.

Abstract: In accordance with a first aspect, a communication device is provided, comprising: a transmitter configured to transmit one or more radio frequency signal pulses to an external communication device; a receiver configured to receive one or more response signals in response to the radio frequency signal pulses transmitted by the transmitter; a signal analyzer configured to detect one or more characteristics of the response signals, to compare the detected characteristics with predefined reference characteristics and to generate an output indicative of a result of comparing the detected characteristics with the predefined reference characteristics; a processing unit configured to determine at least one category to which the external communication device belongs based on the output generated by the signal analyzer. In accordance with a second aspect, a corresponding method of operating a communication device is conceived. In accordance with a third aspect, a corresponding computer program is provided.

Abstract: Various embodiments disclosed herein provides method and system for low latency FPGA based system for inference such as recommendation models. Conventional models for inference have high latency and low throughput in decision making models/processes. The disclosed method and system exploits parallelism in processing of XGB models and hence enables minimum possible latency and maximum possible throughput. Additionally, the disclosed system uses a trained model that is (re)trained using only those features which the model had used during training, remaining features are discarded during retraining of the model. The use of such selected set of features thus leads to reduction in the size of digital circuit significantly for the hardware implementation, thereby greatly enhancing the system performance.

Abstract: A resistive processing unit (RPU) that includes a pair of transistors connected in series providing an update function for a weight of a training methodology to the RPU, and a read transistor for reading the weight of the training methodology. In some embodiments, the resistive processing unit (RPU) further includes a capacitor connecting a gate of the read transistor to the air of transistors providing the update function for the resistive processing unit (RPU). The capacitor stores said weight of training methodology for the RPU.