Abstract: A system and a method of quantizing a pre-trained neural network, includes determining by a layer/channel bit-width determiner for each layer or channel of the pre-trained neural network a minimum quantization noise for the layer or the channel for each master bit-width value in a predetermined set of master bit-width values; and selecting by a bit-width selector for the layer or the channel the master bit-width value having the minimum quantization noise for the layer or the channel. In one embodiment, the minimum quantization noise for the layer or the channel is based on a square of a range of weights for the layer or the channel that is multiplied by a constant to a negative power of a current master bit-width value.

Abstract: A fully connected operation method and a processing device for performing the same are provided. The fully connected operation method designates distribution data and broadcast data. The distribution data is divided into basic data blocks and distributed to parallel processing units, and the broadcast data is broadcasted to the parallel processing units. Operations between the basic data blocks and the broadcasted data are carried out by the parallel processing units before the results are returned to a main unit for further processing. The technical solutions disclosed by the present disclosure provide short Operation time and low energy consumption.

Abstract: An image-processing system that can perform predetermined image processing on a manuscript character in a captured image is provided. A method for image processing includes training a neural network using composite image data including a background image and a manuscript image and correct image data corresponding to the composite image data, obtaining a captured image of an original that contains a manuscript character, and performing predetermined image processing on the captured image using the neural network.

Abstract: Subject matter disclosed herein may relate to storage and/or processing of signals and/or states representative of neural network parameters in a computing device, and may relate more particularly to compressing signals and/or states representative of neural network nodes in a computing device.

Abstract: Systems and techniques for neural network training are described herein, a training set may be received for a neural network. Here, the neural network may comprise a set of nodes arranged in layers and a set of inter-node weights between nodes in the set of nodes. The neural network may then be iteratively trained to create a trained neural network. An iteration of the training may include generating a random unit vector and creating an update vector by calculating a magnitude for the random unit vector based on a degree that the random unit vector matches a gradient—where the gradient is represented by a dual number. The iteration may continue by updating a parameter vector for an inter-node weight by subtracting the update vector from a previous parameter vector of the inter-node weight. The trained neural network may then be used to classify data.

Abstract: Embodiments of the disclosure provide methods and systems for performing machine learning. The method can include: receiving training data; training a machine learning model based on the training data, wherein the machine learning model includes multiple layers each having one or more nodes having one or more connections with a node from another layer of the machine learning model; evaluating weights associated with the connections of the machine learning model, wherein each connection has a corresponding weight; removing, from the machine learning model, one or more connections having a weight that does not satisfy a threshold condition; and after the connections have been removed, updating the machine learning model.

Abstract: A method and apparatus for private information retrieval from a database, wherein the retrieval includes providing a covering vector for a plurality of database entries of the database. The covering vector is defined such that an inner product of the covering vector is zero with more than one member of a covering vector family that includes the covering vector. The retrieval includes generating database queries based on the covering vector and transmitting the database queries to at least two servers. An identical copy of the database may be stored on each of the at least two servers. Shares are received in response to the query, and these shares are aggregated, and a reconstruction algorithm executes to reconstruct the query results.

Abstract: As provided herein, a list of locales of interest in a location may be sorted into one or more categories. A user performing a search for a locale of interest (e.g., a restaurant) may be identified. A local score may be assigned to the locale of interest based upon a number of local users (e.g., users residing in the location) that perform the search. A second user may be determined to be near the locale of interest. A category of interest may be determined for the second user (e.g., an interest in local non-tourist restaurants). Responsive to the category of interest corresponding to the category and the local score of the locale of interest exceeding an interest threshold, the second user may be provided with a recommendation to go to the locale of interest. The locale of interest may be a local favorite restaurant rather than a tourist trap.

Abstract: Embodiments are directed to data verification of business or consumer data. Certain embodiments include a data verification system that receives or selects data to be verified, selects one or more verification methods to verify, update, and/or append/enhance the data. The data verification system may verify the data with one or more data verification methods, either alone or in combination. The methods may include a web-crawling verification method, an agent web verification method, a call verification method, a direct mail method, an email method, an in-person verification method, or other methods. The system has the ability to, automatically or manually, (1) blend automatic and manual segmentation of records or elements by criteria such as industry type, best times of day/month/year to verify, update, or append, cost, and level of importance (2) select the best verification processing method(s), and (3) manage the results and properly verify, update, append/enhance records.

Abstract: In various example embodiments, a system and method for projecting visual aspects into a vector space are presented. A query that includes visual data is received by the system from a client device. A visual aspect indicated in the visual data is analyzed. One or more symbols that correspond to the analyzed visual aspect is generated by the system. The analyzed visual aspect is projected into a vector space using the one or more symbols. A group of projections are identified, the group of projections being within a predetermined distance from the projected visual aspect in the vector space. An interface that depicts the further visual aspects is generated. The interface is displayed on the client device.

Abstract: Provided are a computer program product, system, and method for using a machine learning module to select one of multiple cache eviction algorithms to use to evict a track from the cache. A first cache eviction algorithm determines tracks to evict from the cache. A second cache eviction algorithm determines tracks to evict from the cache, wherein the first and second cache eviction algorithms use different eviction schemes. At least one machine learning module is executed to produce output indicating one of the first cache eviction algorithm and the second cache eviction algorithm to use to select a track to evict from the cache. A track is evicted that is selected by one of the first and second cache eviction algorithms indicated in the output from the at least one machine learning module.

Abstract: Methods, apparatus, systems and articles of manufacture are disclosed to perform malware detection using a generative adversarial network. An example apparatus includes a first encoder network to encode an input sample into a first encoded sample, the first encoder network implemented using a multilayer perception (MLP) network, a generator network to reconstruct the first encoded sample to generate a reconstructed sample, a discriminator network to, in response to obtaining the first encoded sample and the reconstructed sample, generate a loss function based on the reconstructed sample and the input sample, and an optimization processor to, when the loss function satisfies a threshold loss value, classify the input sample as malicious.

Abstract: A method, system, and computer program product for detecting multi-scale objects. The method may include receiving a sample dataset including multi-scale objects associated with a specific environment, where the sample dataset has an existing resolution. The method may also include inputting the sample dataset into a trained neural network, where the trained neural network has a plurality of scale regions. The method may also include processing the sample dataset for detecting the multi-scale objects by means of the trained neural network. The method may also include calculating a distribution of contribution degree in a course of processing the sample dataset, where the contribution degree is associated with each of the plurality of scale regions. The method may also include generating a set of configuration parameters associated with the specific environment for the trained neural network based at least in part on the distribution of contribution degree.

Abstract: The disclosed embodiments relate to a system that determines whether an inferential model is susceptible to spillover false alarms. During operation, the system receives a set of time-series signals from sensors in a monitored system. The system then trains the inferential model using the set of time-series signals. Next, the system tests the inferential model for susceptibility to spillover false alarms by performing the following operations for one signal at a time in the set of time-series signals. First, the system adds degradation to the signal to produce a degraded signal. The system then uses the inferential model to perform prognostic-surveillance operations on the time-series signals with the degraded signal. Finally, the system detects spillover false alarms based on results of the prognostic-surveillance operations.

Abstract: According to one embodiment, a first set of features is received, where each of the features in the first set being associated with a predetermined category. A bloom filter is applied to the first set of features to generate a second set of features. A neural network model is trained by applying the second set of features to a first layer of nodes of the neural network model to generate an output, the neural network model including a plurality of layers of nodes coupled to each other via a connection. The output of the neural network model is compared with a target value associated with the predetermined category to determine whether the neural network model satisfies a predetermined condition.

Abstract: Disclosed herein are methods and system for training artificial intelligence models configured to execute image segmentation techniques. The methods and system describe a server that receives a first image including a set of pixels depicting multiple objects. The server also receives a second image having a second set of pixels depicting the same set of objects. The server then analyzes the pixels from the first and second images. When a difference between at least one visual attribute of a pixel within the second image and a corresponding pixel within the first image satisfies a predetermined threshold, it will be encoded as spikes to send to the model, the model will be trained using supervised STDP rule by revising weights associated with the nodes within the AI model where the node corresponds to the pixels within the first and/or the second image.

Abstract: Some embodiments include a system operable to construct hierarchical training data sets for use with machine-learning for multiple controlled devices. Other embodiments of related systems and methods are also provided.

Abstract: A method of quantizing an artificial neural network includes dividing an input distribution of the artificial neural network into a plurality of segments, generating an approximated density function by approximating each of the plurality of segments, calculating at least one quantization error corresponding to at least one step size for quantizing the artificial neural network, based on the approximated density function, and determining a final step size for quantizing the artificial neural network based on the at least one quantization error.

Abstract: Subject matter disclosed herein may relate to storage of signals and/or states representative of parameters in a computing device, and may relate more particularly to storage of signals and/or states representative of neural network parameters in a computing device.

Abstract: A scalable stream synaptic supercomputer for extreme throughput neural networks is provided. The firing state of a plurality of neurons of a first neurosynaptic core is determined substantially in parallel. The firing state of the plurality of neurons is delivered to at least one additional neurosynaptic core substantially in parallel.

Abstract: A computer-based system and method for determining similarity between at least two heterogenous unstructured data records and for optimizing processing performance. A plurality of occupational data records is generated and, for each of the occupational data records, a respective vector is created to represent the occupational data record. Each of the vectors is sliced into a plurality of chunks. Thereafter, semantic matching of the chunks occurs in parallel, to compare at least one occupational data record to at least one other occupational data record simultaneously and substantially in real time. Thereafter, values representing similarities between at least two of the occupational data records are output.

Abstract: A memory component can include memory cells with a memory region to store a machine learning model and input data and another memory region to store host data from a host system. The memory component can include an in-memory logic, coupled to the memory cells, to perform a machine learning operation by applying the machine learning model to the input data to generate an output data. A bus can receive additional data from the host system and can provide the additional data to the other memory region or the in-memory logic based on a characteristic of the additional data.

Abstract: A method for operating artificial neural network with nonvolatile memory device having at least one artificial neural nonvolatile memory network. The forgoing artificial neural nonvolatile memory network (ANN) comprises M×N numbers nonvolatile memory cells that are arranged to form a memory array, and the nonvolatile memory cell can be a non-overlapped implementation (NOI) MOSFET, a RRAM element, a PCM element, a MRAM element, or a SONOS element. By applying this novel method to the ANN, it is able to perform feedforward and recurrent operations in the M×N numbers of nonvolatile memory devices in the ANN, so as to adjust or correct the weights stored in the M×N numbers of nonvolatile memory devices.

Abstract: Providing knowledge representation of material content being consumed by a user combines the user's current behavioral data and data from external sources such as internet web sites and social media network. Visual representations of entities and their relationships in the content being consumed by the user are created while the user is consuming content, and displayed via a graphical user interface.

Abstract: A computing node in a distributed storage system is configured to send an instruction of migration of a parity fragment of a plurality of data fragments from a first-level storage medium to a second-level storage medium, where performance of the second-level storage medium is lower than the first-level storage medium.

Abstract: Methods, systems, and computer-readable media for updating a machine learning model utilized in a search engine operation. The method identifies a set of search queries in stored search queries corresponding to a service and apply the identified set of search queries to the search engine to generate one or more search results for the service. Each search result has an assigned aggregate based on values of a set of parameters associated with the service. The method then analyzes the values of the set of parameters to determine a tradeoff point of each parameter to determine one or more weights to apply to the machine learning model based on the tradeoff points. The method stores the determined one or more weights and applies them to the machine learning model for a search query corresponding to the service.

Abstract: Method, electronic device, and computer readable medium embodiments are disclosed. In one embodiment, a method includes receiving image data, manipulating the received image data based on a set of transform parameters, and analyzing the manipulated image data to generate metadata. The metadata statistically describes the received image data. The method also includes selecting a neural network from a plurality of neural networks to perform a second analysis, wherein the neural network is selected based on the generated metadata. The method additionally includes performing a second analysis of the received image data by the selected neural network based on the generated metadata to extract information from the received image data.

Abstract: Methods and systems for receiving a request to implement a neural network comprising an average pooling layer on a hardware circuit, and in response, generating instructions that when executed by the hardware circuit, cause the hardware circuit to, during processing of a network input by the neural network, generate a layer output tensor that is equivalent to an output of the average pooling neural network layer by performing a convolution of an input tensor to the average pooling neural network layer and a kernel with a size equal to a window of the average pooling neural network layer and composed of elements that are each an identity matrix to generate a first tensor, and performing operations to cause each element of the first tensor to be divided by a number of elements in the window of the average pooling neural network layer to generate an initial output tensor.

Abstract: In a multifunction peripheral according to the present invention, a sheet selection screen for selecting a sheet that is the output destination is presented. A “check sheet size” button is provided on the sheet selection screen. When the “check sheet size” button is operated, an optimal sheet size identification function is started, and a document size identification process including pre-scan is performed. Accordingly, the size of a document is identified. Further, an optimal sheet size identification process is performed. Thus, the optimal sheet size, which is a suitable sheet size corresponding to the size of the document, is identified.

Abstract: A device for extracting dynamic information comprises a convolutional neural network, wherein the device is configured to receive a sequence of data blocks acquired over time, each of said data blocks comprising a multi-dimensional representation of a scene. The convolutional neural network is configured to receive the sequence as input and to output dynamic information on the scene in response, wherein the convolutional neural network comprises a plurality of modules, and wherein each of said modules is configured to carry out a specific processing task for extracting the dynamic information.

Abstract: A Deep Neural Networks (DNN) analysis method, system, and computer program product include characterizing a space of possible configurations for a DNN, evaluating a metric-of-interest for a configuration of the possible configurations, and searching the space to identify a configuration of the possible configurations that maximizes the metric-of-interest.

Abstract: Segment valuation techniques usable in a digital medium environment are described. To do so, a segment valuation system first identifies the attributes that are significant in achievement of a desired metric (e.g., conversion) and then values segments based on those significant attributes. Attributes are selected from the trained model based on significance of those attributes towards achieving the desired metric. A valuation of a segment may then be calculated based on the valuations of these attributes. For example, inclusion of the selected attributes within a segment, and the valuations of those selected attributes, is then used by the segment valuation system to generate data describing a value of the segment towards achieving the metric.

Abstract: Provided are a neural network system, an application processor including the same, and a method of operating the neural network system. The neural network system includes an operation allocating manager configured to divide a request including a plurality of operations for neural network processing into a plurality of sub-requests each including at least one operation, allocate the plurality of sub-requests to dedicated hardware and one or more pieces of general-purpose hardware, and allocate memories for input and output for each of the plurality of sub-requests based on cost information indicating respective costs needed for hardware to access the memories; and a neural network device configured to execute the sub-requests by using the dedicated hardware and the one or more pieces of general-purpose hardware according to a result of the operation allocation. The operation allocating manager allocates different memories to at least two sub-requests of the plurality of sub-requests.

Abstract: A sample is obtained from a test sample set. The sample is input into a plurality of models included in a model set that are to be tested, where the plurality of models include at least one neural network model. A plurality of output results are obtained, including obtaining, from each model of the plurality of models, a respective output result. A test result is determined based on the plurality of output results, where the test result includes at least one of a first test result or a second test result, where the first test result includes a plurality of output result accuracies. In response to determining that the test result does not satisfy a predetermined condition, a new sample is generated based on the sample and a predetermined rule, and the new sample is added to the test sample set.

Abstract: Embodiments of the invention relate to canonical spiking neurons for spatiotemporal associative memory. An aspect of the invention provides a spatiotemporal associative memory including a plurality of electronic neurons having a layered neural net relationship with directional synaptic connectivity. The plurality of electronic neurons configured to detect the presence of a spatiotemporal pattern in a real-time data stream, and extract the spatiotemporal pattern. The plurality of electronic neurons are further configured to, based on learning rules, store the spatiotemporal pattern in the plurality of electronic neurons, and upon being presented with a version of the spatiotemporal pattern, retrieve the stored spatiotemporal pattern.

Abstract: The subject technology receives a neural network model in a model format, the model format including information for a set of layers of the neural network model, each layer of the set of layers including a set of respective operations. The subject technology generates neural network (NN) code from the neural network model, the NN code being in a programming language distinct from the model format, and the NN code comprising a respective memory allocation for each respective layer of the set of layers of the neural network model, where the generating comprises determining the respective memory allocation for each respective layer based at least in part on a resource constraint of a target device. The subject technology compiles the NN code into a binary format. The subject technology generates a package for deploying the compiled NN code on the target device.

Abstract: There is provided a computer implemented method for detection of likelihood of malignancy in an anatomical image of a patient for treatment planning, comprising: receiving an anatomical image, feeding the anatomical image into a global component of a model trained to output a global classification label, feeding the anatomical image into a local component of the model trained to output a localized boundary, feeding the anatomical image patch-wise into a patch component of the model trained to output a patch level classification label, extracting a respective set of regions of interest (ROIs) from each one of the components, each ROI indicative of a region of the anatomical image likely to include an indication of malignancy, aggregating the ROIs from each one of the components into an aggregated set of ROIs, and feeding the aggregated set of ROIs into an output component that outputs an indication of likelihood of malignancy.

Abstract: Systems and methods are described for automatically assigning roles for an incident to user profiles using machine learning techniques to optimize the overall suitability of the automatic role assignments and to adapt to manual role reassignments. One or more roles of a plurality of roles for an incident are automatically assigned to each user profile of a plurality of user profiles by the computer-based system applying a combinatorial optimization routine. An indication is displayed on a screen of a determined relative suitability of a first user profile for each role of the plurality of roles. The computer-based system is configured to receive a manual reassignment selection of a different role of the plurality of roles to be assigned to the first user profile and the combinatorial optimization routine is updated based on the manual reassignment selection.

Abstract: A lower-dimensional representation (e.g., approximation) of a dataset is determined. The lower-dimensional representation can be used, for example, to perform semantic document analysis. Given a matrix of input data points, where each entry of the matrix indicates a number of times a particular term in a set of terms appears in a particular document in a set of documents, a lower-dimensional compressed matrix is obtained from the matrix by sampling rows of the matrix based on a target rank parameter, a desired accuracy tolerance, leverage scores calculated for the rows, and/or distances from rows of the matrix to a span of the initial set of sampled rows. The compressed matrix is used to determine a similarity metric indicative of a degree of similarity between documents. The documents can then be classified into a same document cluster or different clusters based on whether the similarity metric satisfied a threshold value.

Abstract: One embodiment provides a method for relevant language-independent terminology extraction from content, the method including extracting lexicon items from the content based on context extraction patterns using statistical processing. Feedback on the extracted lexicon items is received to automatically tune scores and thresholds for the context extraction patterns. Available Linked Data is leveraged for a bootstrap source. The relevant language-independent terminology extraction is bootstrapped using the bootstrap source.

Abstract: A system including a confidence assessment module that implements a neural network to assess the likelihood that codes associated with a patient's encounter with a healthcare organization are accurate. The confidence assessment module may also be incrementally trained.

Abstract: A system and method for controlling a nodal network. The method includes estimating an effect on the objective caused by the existence or non-existence of a direct connection between a pair of nodes and changing a structure of the nodal network based at least in part on the estimate of the effect. A nodal network includes a strict partially ordered set, a weighted directed acyclic graph, an artificial neural network, and/or a layered feed-forward neural network.

Abstract: Systems and methods for expert-assisted classification are described herein. An example method for evaluating an expert-assisted classifier can include providing a cascade classifier including a plurality of classifier stages; and providing a simulated expert stage between at least two of the classifier stages. The simulated expert stage can be configured to validate or contradict an output of one of the at least two classifier stages. The method can also include classifying each of a plurality of records into one of a plurality of categories using the cascade classifier combined with the simulated expert stage; and determining whether the simulated expert stage improves performance of the cascade classifier.

Abstract: A method for automatically generating a neural network architecture includes: loading a first computer-readable representation of a first neural network architecture including first genes representing parameters of the first neural network architecture; generating a first neural network including neurons in accordance with the first genes; deploying the first neural network in a robotic controller; training the first neural network by supplying inputs to an input processing unit connected to the first neural network and receiving outputs from an output processing unit connected to the first neural network, the training including updating synaptic weights of connections between the neurons based on responses to the inputs supplied to the input processing unit; evaluating a performance of the first neural network architecture; and generating, by the computer system, an updated computer-readable representation of an updated neural network architecture based on the evaluation of the performance the first neural n

Abstract: A transmitting apparatus is provided. The transmitting apparatus includes: an encoder configured to perform a low-density parity check (LDPC) encoding on input bits using a parity check matrix to generate an LDPC codeword comprising information word bits and parity bits; an interleaver configured to interleave the LDPC codeword; and a modulator configured to map the interleaved LDPC codeword onto a modulation symbol, wherein the modulator is further configured to map a bit included in a predetermined bit group from among a plurality of bit groups constituting the LDPC codeword onto a predetermined bit of the modulation symbol.

Abstract: Improved systems and methods for automated machine-learning, zero-day malware detection. Embodiments include a system and method for detecting malware using multi-stage file-typing and, optionally pre-processing, with fall-through options.

Abstract: Methods and apparatuses that apply a hierarchical-decomposition reinforcement learning technique to train one or more AI objects as concept nodes composed in a hierarchical graph incorporated into an AI model. The individual sub-tasks of a decomposed task may correspond to its own concept node in the hierarchical graph and are initially trained on how to complete their individual sub-task and then trained on how the all of the individual sub-tasks need to interact with each other in the complex task in order to deliver an end solution to the complex task. Next, during the training, using reward functions focused for solving each individual sub-task and then a separate one or more reward functions focused for solving the end solution of the complex task. In addition, where reasonably possible, conducting the training of the AI objects corresponding to the individual sub-tasks in the complex task, in parallel at the same time.

Abstract: Methods and apparatus are provided for evaluating combinatorial processes using simulation techniques and multiple parallel statistical analyses of real-world data. A simulation model is generated that simulates one or more steps of a combinatorial process. The simulation model comprises key features of the combinatorial process. A plurality of first data mining tasks are performed in parallel over real data of the combinatorial process to obtain key feature prediction models that estimate the key features. The key feature prediction models are bound to the simulation model. Query types to be supported are identified and a plurality of simulation runs are generated in parallel, comprising simulated data for the supported query types. A plurality of second data mining tasks are performed in parallel over the plurality of simulation runs to build global prediction models to answer queries of each supported query type. An answer to a user query is determined using the global prediction models.

Abstract: Resolving ambiguities in regulatory documents is necessary to ensure organizations and people are able to be best possible compliant with regulations or standards. Current approaches attempting to automatically resolve ambiguities in regulatory documents have limitations when it comes to incorporating fairness or reduce chances of subjective interpretation. Embodiments of the present disclosure provide a method and system for automatically resolving ambiguities in regulations. To disambiguate a given regulatory sentence the method augments the regulation sentence with relevant internal information extracted using a set of predefined linguistic patterns and relevant external information extracted from external sources identified using a Neural Network (NN) model.

Abstract: Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for generating an output sequence from an input sequence. In one aspect, one of the systems includes an encoder neural network configured to receive the input sequence and generate encoded representations of the network inputs, the encoder neural network comprising a sequence of one or more encoder subnetworks, each encoder subnetwork configured to receive a respective encoder subnetwork input for each of the input positions and to generate a respective subnetwork output for each of the input positions, and each encoder subnetwork comprising: an encoder self-attention sub-layer that is configured to receive the subnetwork input for each of the input positions and, for each particular input position in the input order: apply an attention mechanism over the encoder subnetwork inputs using one or more queries derived from the encoder subnetwork input at the particular input position.