Abstract: This disclosure relates to methods, non-transitory computer readable media, and systems that can classify term sequences within a source text based on textual features analyzed by both an implicit-class-recognition model and an explicit-class-recognition model. For example, by applying machine-learning models for both implicit and explicit class recognition, the disclosed systems can determine a class corresponding to a particular term sequence within a source text and identify the particular term sequence reflecting the class. The dual-model architecture can equip the disclosed systems to apply (i) the implicit-class-recognition model to recognize implicit references to a class in source texts and (ii) the explicit-class-recognition model to recognize explicit references to the same class in source texts.

Abstract: Described herein are embodiments for end-to-end reinforcement learning based coreference resolution models to directly optimize coreference evaluation metrics. Embodiments of a reinforced policy gradient model are disclosed to incorporate reward associated with a sequence of coreference linking actions. Furthermore, maximum entropy regularization may be used for adequate exploration to prevent a model embodiment from prematurely converging to a bad local optimum. Experiments on datasets compared with state-of-the-art methods verified the effectiveness of embodiments.

Abstract: A keyword extraction method includes: extracting candidate words from an original document to form a first word set; acquiring a first association degree between each first word thereof and the original document, and determining a second word set according to the first association degree; for each second word in the second word set, inquiring, in a word association topology, at least one node word satisfying a condition of association with the second word and forming a third word set, the word association topology indicating an association relation among multiple node words in a predetermined field; and determining a union set of the second and third word sets, acquiring a second association degree between each candidate keyword in the union set and the original document, and selecting, according to the second association degree, at least one candidate keyword from the union set, to form a keyword set of the original document.

Abstract: A framework for enhanced search is provided to enable the user to contextually search documents, based on specified events. A management framework provides systems for obtaining a user's contextual data and the user's interaction data for which a Context Graph and an Interaction Graph are created, respectively. The user's context data include the various events associated with the user. The user's interaction data include times and locations the user interacted with a particular document, which device was used, the type of interaction, etc. The data from the two graphs are correlated and joined into an event graph. Upon receiving a natural language query from a user, one or more contextual query elements are identified, and the event graph is spanned to identify matching results which are then provided to the requesting user device.

Abstract: Techniques for automatically extracting data from documents using multiple deep learning models are provided. According to one set of embodiments, a computer system can receive a document in an electronic format and can segment, using an image segmentation deep learning model, the document into a plurality of segments, where each segment corresponds to a visually discrete portion of the document and is classified as being one of a plurality of types. The computer system can then, for each segment in the plurality of segments, retrieve text in the segment using optical character recognition (OCR) and extract data in the segment from the retrieved text using a named entity recognition (NER) deep learning model, where the retrieving and the extracting are performed in a manner that takes into account the segment's type.

Abstract: A natural language processing method includes obtaining a to-be-processed phrase, where the to-be-processed phrase includes M words, determining polarity characteristic information of m to-be-processed words in the M words, where polarity characteristic information of an ith word in the m to-be-processed words includes n polarity characteristic values, and each polarity characteristic value corresponds to one sentiment polarity, determining a polarity characteristic vector of the to-be-processed phrase based on the polarity characteristic information of the m to-be-processed words, where the polarity characteristic vector includes n groups of components in a one-to-one correspondence with n sentiment polarities, and determining a sentiment polarity of the to-be-processed phrase based on the polarity characteristic vector of the to-be-processed phrase using a preset classifier, and outputting the sentiment polarity.

Abstract: The disclosure herein describes determining topics of communication transcripts using trained summarization models. A first communication transcript associated with a first communication is obtained and divided into a first set of communication segments. A first set of topic descriptions is generated based on the first set of communication segments by analyzing each communication segment of the first set of communication segments with a generative language model. A summarization model is trained using the first set of communication segments and associated first set of topic descriptions as training data. The trained summarization model is then applied to a second communication transcript and, based on applying the trained summarization model to the second communication transcript, a second set of topic descriptions of the second communication transcript is generated.

Abstract: A vision-based system for identifying textual sentiment present in a text source image, comprising a camera module adapted to capture the text source image, and a vision-based text analyzing device adapted to analyze the input image, identify a text input string, and determine the sentiment embodied by the text input string, the vision-based text analyzing device is further adapted to revise the text input string to embody an alternate sentiment, the vision-based system has an output display for presenting sentiment classification results and the revised text string via in an augmented reality format.

Abstract: A method for classifying a digital asset, comprising: retrieving from a repository, according to an initial identifier, a digital asset; computing at least one asset score, each indicative of a degree of confidence that the digital asset has a respective offense classification selected from a plurality of offense classifications; subject to the at least one asset score being less than at least one threshold score: adding to a list of identifiers at least one identifier extracted from the digital asset; and in each of at least one iteration: retrieving from the repository at least one other digital asset, according to at least one selected identifier selected from the list of identifiers; computing a plurality of other asset scores, each associated with one other digital asset of the at least one other digital asset and indicative of another degree of confidence that the other digital asset has another respective offense classification.

Abstract: A device includes a memory adapted to store a list in a file or database comprising a plurality of vocabulary words in a first language and, for each vocabulary word, a corresponding word in a second language, a display device, and a processor. The processor is adapted to receive a plurality of words in the first language, select one or more words among the plurality of words, based on one or more predetermined criteria, translate, match or equate the one or more selected words from the first language to words of the second language, and cause the display device to display the plurality of words, wherein one or more first words that are in the plurality of words and are not among the one or more selected words which are displayed in the first language and one or more second words that are in the plurality of words and are among the one or more selected words are displayed in the second language.

Abstract: This invention provides an aimer assembly for a vision system that is coaxial (on-axis) with the camera optical axis, thus providing an aligned aim point at a wide range of working distances. The aimer includes a projecting light element located aside the camera optical axis. The beam and received light from the imaged (illuminated) scene are selectively reflected or transmitted through a dichoric mirror assembly in a manner that permits the beam to be aligned with the optical axis and projected to the scene while only light from the scene is received by the sensor. The aimer beam and illuminator employ differing light wavelengths. In a further embodiment, an internal illuminator includes a plurality of light sources below the camera optical axis. Some of the light sources are covered by a prismatic structure for close distance, and other light sources are collimated, projecting over a longer distance.

Abstract: A detecting system includes: a sensor (10) that includes an antenna unit (11) formed with a metal pattern, and a back surface reflector (13) that faces the antenna unit (11) via an isolation layer (12); and a reader (20) that transmits electromagnetic waves (Fa) to the sensor (10), receives reflected waves (Fr) from the sensor (10), and compares the reflection characteristics of the sensor (10) detected from the reflected waves (Fr) with the reflection characteristics of the sensor (10) stored in advance, to detect a state change in the sensor (10).

Abstract: A gaming table contains a number of antennas that are de-tuned from the resonant frequency of a single RFID tag. This increases the power required to read the single tag but matches the resonant frequency for reading a stack of RFID tags. The gaming table may further include a network analyzer and a set of capacitors that are dynamically switched among the antennas according to measuring the reflection coefficient of the antennas.

Abstract: A reader system for reading information on at least one identification tag associated with a syringe connected to a fluid injector has a detector and a background screen spaced apart from the detector such that at least one of the at least one identification tags is positioned in an optical path between the detector and the background screen when the syringe is connected to the fluid injector. The reader system further has an illumination system configured for illuminating at least a portion of the background screen. The detector is configured for detecting and decoding the at least one identification tag. Methods for reading information on at least one identification tag associated with a syringe are also described.

Abstract: Barcode readers and methods for directing fields-of-view of imaging sensors of barcode readers are disclosed herein. An example barcode reader includes a housing, an imaging sensor positioned within the housing, and a transflective mirror positioned within the housing and in a path of a field-of-view of the imaging sensor. The transflective mirror reflects at least a first portion of the field-of-view of the imaging sensor in a first direction with the transflective mirror in a reflective state and allows the field-of-view of the imaging sensor to pass through and continue in a second direction, different than the first direction, with the transflective mirror in a transmissive state.

Abstract: A bioptic scanner optical arrangement with a single sensor split four ways is disclosed herein. An example bioptic scanner optical arrangement includes a housing, an imaging assembly having a primary FOV, a decode module, a generally horizontal window supported by the housing, a generally upright window supported by the housing, and a mirror arrangement positioned within the interior region. The mirror arrangement is configured to divide the primary FOV into a plurality of subfields, to redirect at least two of the plurality of subfields through the generally horizontal window, and to redirect at least another two of the plurality of subfields through the generally upright window. The bioptic barcode reader has no other imaging assembly communicatively coupled to the decode module and used to process images for decoding indicia.

Abstract: An article of manufacture may include a matrix barcode on a physical medium and associated with an environment. The matrix barcode may include a plurality of colors, where at least one of the plurality of colors is a least prevalent color, of a plurality of prevalent colors in the environment.

Abstract: A method for identity (ID) verification, including: acquiring one or more ID indications of at least one item using a hand held verifying device; automatically verifying by the verifying device a desired relation between the one or more acquired ID indications and stored verification data associated with the verifying device; physically marking the item using the verifying device with a verifying marking if the desired relation is automatically verified.

Abstract: The present invention relates to a method of reading out information from a data carrier and to a data carrier utilizing the concept of structured-illumination microscopy or saturated structured-illumination microscopy.

Abstract: Software performance can be predicted based on different system configurations. In one example, a computing device can receive historical datasets associated with copies of a software application executed by a group of computing environments during a prior timespan. Each historical dataset can indicate respective changes during the prior timespan to at least one performance characteristic of one of the copies of the software application executed by one of the computing environments in the group. Each computing environment in the group can being configured differently than the other computing environments in the group. The computing device can also convert the historical datasets into training data for a machine-learning model, and train the machine-learning model. This can yield a trained machine-learning model configured to generate a forecast of the performance characteristic for the software application over a future timespan.

Abstract: An assembly change detection method based on attention mechanism, including: establishing a three-dimensional model of an assembly body, adding a tag to each part in the three-dimensional model, setting several assembly nodes, obtaining depth images of the three-dimensional model under each assembly node in different viewing angles, and obtaining a change tag image of a added part at each assembly node; selecting two depth images at front and back moments in different viewing angles as training samples; performing semantic fusion, feature extraction, attention mechanism processing and metric learning sequentially on the training samples, training a detection model, continuously selecting training samples to train the detection model, saving model parameters with optimal similarity during training, completing training; and obtaining depth images of successive assembly nodes during assembling the assembly body, inputting depth images into trained detection model, and outputting change image of added part of the

Abstract: A computing device accesses a machine learning model trained on training data of first bonding operations (e.g., a ball and/or stitch bond). The first bonding operations comprise operations to bond a first set of multiple wires to a first set of surfaces. The machine learning model is trained by supervised learning. The device receives input data indicating process data generated from measurements of second bonding operations. The second bonding operations comprise operations to bond a second set of multiple wires to a second set of surfaces. The device weights the input data according to the machine learning model. The device generates an anomaly predictor indicating a risk for an anomaly occurrence in the second bonding operations based on weighting the input data according to the machine learning model. The device outputs the anomaly predictor to control the second bonding operations.

Abstract: Systems, methods, devices, server computers, storage media, and instructions for prioritized device action triggered by device scan data are described. In one embodiment, a mobile device performs a method that involves executing a messaging application with an image capture interface and a scanning input. An associated scanning mode comprises capture of scan data from a plurality of input/output modules of the first client device, analyzes the scan data to identify one or more scan data patterns by matching at least a portion of the scan data against a set of data patterns, and selects a priority system action based on the results of the matching of the portion of the scan data against the set of data patterns. In some embodiments, the priority system action is selected based on a priority ranking for identified scan data types.

Abstract: A printing system is disclosed.

Abstract: An information product or item, such as a for sale sign, can be configured with contact information and/or elements that link the information product to a networked system and related services (with the product being called a for information product or FIP in the description). These services can make available applications such as phone and web application that can be helpful in providing an additional mechanism/electronic media as it relates to the FIP and/or any FIP related product/object/item. These services may aid someone to whom the FIP belongs in fielding, FIP related, viewer inquires while providing the FIP viewer with more options, upon making requests to said system services. The FIP may make available linked application system services to viewers of the FIP, which may belong to someone who does not know how to link such services in other ways.

Abstract: Instruction data is configured to include mode information indicating a print mode in download printing. The print mode includes a first mode and a second mode different from the first mode. The download printing includes: in response to extracting the mode information indicating the first mode, performing the download printing by the first mode, the first mode being a mode of, after completing reception of an entirety of a document, starting printing based on the entirety of the document; and in response to extracting the mode information indicating the second mode, performing the download printing by the second mode, the second mode being a mode of, before completing reception of the entirety of the document, starting printing based on a received portion of the document. The second mode is a mode in which reception of the document and printing based on the document are performed in parallel.

Abstract: The invention relates to a chip card with a dual, contactless and contact-based communication interface, comprising a card body made up of two metal sheets and an electronic module provided with a terminal block and an antenna designed for radio-frequency communication with the antenna of a remote reader. The electronic module includes an active communication circuit powered by a battery in the chip card, and configured to generate a carrier signal capable of interacting with the carrier signal emitted by a remote contactless reader in order to establish active radio-frequency communication between the chip card and the remote reader.

Abstract: A self-tuning RFID device having an input capacitance that is adjustable in response to a detected signal. The self-tuning RFID device preferably comprises a variable capacitance RFID chip coupled to an inductor, and an input circuit driven by the detected signal from the variable capacitance RFID chip. A change in capacitance with the detected signal is delayed by a specific amount of time, thereby allowing the self-tuning RFID device to function as a dual mode EAS and RFID tag. The ESA functionality can be deactivated by a high field at or near its resonance frequency without disabling the RFID functionality. The effect of the high field may change the input capacitance permanently changing its resonance.

Abstract: Discloses a combined ultra-wideband cross-polarized chipless RFID tag based on MFCC feature coding, which comprises a tag patch unit, a dielectric substrate and a grounding layer, wherein the tag patch unit comprises a barcode-type resonant unit and a double L-type resonant unit; the barcode-type resonant unit consists of five identical rectangular patches arranged in parallel and rotated counterclockwise; the double L-type resonant unit is formed by reversely combining two L-type patches composed of four identical rectangular patches; a transmitting antenna transmits horizontally polarized electromagnetic waves as interrogation signals, the scattered waves reflected by the tag are acquired by a receiving antenna, a receiver acquires the spectrum of the scattered waves to convert the spectrum into time domain signals by inverse Fourier transform, the response of the tag is extracted through a window, and MFCC features are extracted by pre-emphasis and short-time Fourier transform.

Abstract: Connection bridges (CBR) for dual-interface transponder chip modules (TCM) 200 may have an area which is substantially equal to or greater than an area of a contact pad (CP) of a contact pad array (CPA). A given connection bridge may be L-shaped and may comprise (i) a first portion disposed external to the contact pad array and extending parallel to the insertion direction, and (ii) a second portion extending from an end of the first portion perpendicular to the insertion direction to within the contact pad array (CPA) such as between C1 and C5. The connection bridge may extend around a corner of the contact pad array, may be large enough to accommodate wire bonding, and may be integral with a coupling frame (CF) extending around the contact pad array. The transponder chip modules may be integrated into a smart card (SC).

Abstract: Aspects of the present disclosure address systems and methods for fixed-point quantization using a dynamic quantization level adjustment scheme. Consistent with some embodiments, a method comprises accessing a neural network comprising floating-point representations of filter weights corresponding to one or more convolution layers. The method further includes determining a peak value of interest from the filter weights and determining a quantization level for the filter weights based on a number of bits in a quantization scheme. The method further includes dynamically adjusting the quantization level based on one or more constraints. The method further includes determining a quantization scale of the filter weights based on the peak value of interest and the adjusted quantization level. The method further includes quantizing the floating-point representations of the filter weights using the quantization scale to generate fixed-point representations of the filter weights.

Abstract: A method for generating an executable program to run on a system of one or more processor chips each comprising a plurality of tiles. The method comprises: receiving a graph comprising a plurality of data nodes, compute vertices and directional edges, wherein the graph is received in a first graph format that does not specify which data nodes and vertices are allocated to which of the tiles; and generating an application programming interface, API, for converting the graph, to determine a tile-mapping allocating the data nodes and vertices amongst the tiles. The generating of the API comprises searching the graph to identify compute vertices which match any of a predetermined set of one or more compute vertex types. The API is then called to convert the graph to a second graph format that includes the tile-mapping, including the allocation by the assigned memory allocation functions.

Abstract: Proposed are a method and apparatus for accelerating data processing in a neural network. The apparatus for accelerating data processing in a neural network may include: a control unit configured to quantize data by at least one method according to a characteristic of data calculated at a node forming at least one layer constituting the neural network, and to separately perform calculation at the node according to the quantized data; and memory configured to store the quantized data.

Abstract: An image analysis method and an image analysis system are disclosed. The method may include extracting training graphic data including at least one first node corresponding to a plurality of histological features of a training tissue slide image, and at least one first edge defined by a relationship between the histological features The method may also include determining a parameter of a readout function by training a graph neural network (GNN) using the training graphic data and training output data corresponding to the training graphic data. The method may also include extracting inference graphic data including at least one second node corresponding to a plurality of histological features of an inference tissue slide image, and at least one second edge decided by a relationship between the histological features, The method may further include deriving inference output data by the readout function after inputting the inference graphic data to the GNN.

Abstract: A method for generating an executable program to run on a system of one or more processor chips each comprising a plurality of tiles. The method comprises: receiving a graph comprising a plurality of data nodes, compute vertices and directional edges, wherein the graph is received in a first graph format that does not specify which data nodes and vertices are allocated to which of the tiles; and generating an application programming interface, API, for converting the graph, to determine a tile-mapping allocating the data nodes and vertices amongst the tiles. The generating of the API comprises searching the graph to identify compute vertices which match any of a predetermined set of one or more compute vertex types. The API is then called to convert the graph to a second graph format that includes the tile-mapping, including the allocation by the assigned memory allocation functions.

Abstract: Technologies for a neural belief reasoner model generative models that specifies belief functions are described. Aspects include receiving, by a device operatively coupled to a processor, a request for a belief function, and processing, by the device, the request for the belief function in the generative model based on trained probability parameters and a minimization function to determine a generalized belief function defined by fuzzy sets. Data corresponding to the generalized belief function is output, e.g., as a belief value and plausibility value.

Abstract: A computer-implemented method, a computer program product, and a computer system for multi-sample dropout in deep neural network training. A computer creates multiple dropout samples in a minibatch, starting from a dropout layer and ending at a loss function layer in a deep neural network. At the dropout layer in the deep neural network, the computer applies multiple random masks for respective ones of the multiple dropout samples. At a fully connected layer in the deep neural network, the computer applies a shared parameter for all of the multiple dropout samples. After the loss function layer in the deep neural network, the computer calculates a final loss value, by averaging loss values of the respective ones of the multiple dropout samples.

Abstract: A computing device receives a data X and Y, each having N samples. A function f(x,y) is defined to be a trainable neural network based on the data X and the data Y. A permuted version of the data Y is created. A loss mean is computed based on the trainable neural network f(x,y), the permuted version of the sample data Y, and a trainable scalar variable ?. A loss with respect to the scalar variable ? and the trainable neural network is minimized. Upon determining that the loss is at or below the predetermined threshold, estimating a mutual information (MI) between a test data XT and YT. If the estimated MI is above a predetermined threshold, the test data XT and YT is deemed to be dependent. Otherwise, it is deemed to be independent.

Abstract: The present disclosure provides systems, methods and computer-readable media for optimizing the neural architecture search for the automated machine learning process. In one aspect, neural architecture search method including selecting a neural architecture for training as part of an automated machine learning process; collecting statistical parameters on individual nodes of the neural architecture during the training; determining, based on the statistical parameters, active nodes of the neural architecture to form a candidate neural architecture; and validating the candidate neural architecture to produce a trained neural architecture to be used in implemented an application or a service.

Abstract: Embodiments relate to a neural processor that includes one or more neural engine circuits and planar engine circuits. The neural engine circuits can perform convolution operations of input data with one or more kernels to generate outputs. The planar engine circuit is coupled to the plurality of neural engine circuits. A planar engine circuit can be configured to multiple modes. In an elementwise mode, the planar engine circuit may combine two tensors by performing operations element by element. The planar engine circuit may support elementwise operation for two tensors that are in different sizes and ranks. The planar engine circuit may perform a broadcasting operation to duplicate one or more values across one or more channels to make a smaller tensor matching the size of the larger tensor.

Abstract: Disclosed herein are a neural network generation method for neuromorphic computing and an apparatus for the same. The neural network generation method for neuromorphic computing includes selecting at least one part of a brain corresponding to a neural network function requested to be generated by an application, determining whether an existing neural network corresponding to the at least one part of the brain is present in a neural network database, when it is determined that no existing neural network is present in the neural network database, generating new neural network configuration information corresponding to the part of the brain based on a brain information database, generating a new neural network by mapping the new neural network configuration information to neuromorphic hardware based on the new neural network configuration information, and storing the new neural network in the neural network database.

Abstract: An artificial neuron for a neuromorphic chip comprises a synapse with resistive memory representative of a synaptic weight. The artificial neuron comprises a read circuit, an integration circuit and a logic circuit interposed between the read circuit and the integration circuit. The read circuit is configured to impose on the synapse a read voltage independent of the membrane voltage and to provide an analogue value representative of the synaptic weight. The logic circuit is configured to generate from the analogue value a pulse having a duration. The integration circuit comprises an accumulator of synaptic weights at the terminals of which a membrane voltage is established and a comparator configured to emit a postsynaptic pulse if a threshold is exceeded by the membrane voltage. Moreover, it comprises a source of current controlled by the pulse to inject a current into the accumulator of synaptic weights during this duration.

Abstract: A method of training a neural network includes, at a local computing node, receiving remote parameters from a set of one or more remote computing nodes, initiating execution of a forward pass in a local neural network in the local computing node to determine a final output based on the remote parameters, initiating execution of a backward pass in the local neural network to determine updated parameters for the local neural network, and prior to completion of the backward pass, transmitting a subset of the updated parameters to the set of remote computing nodes.

Abstract: The user is to be informed of the reliability of the machine-learned model based on the current input relative to the training data used to train the model or the model itself. In a medical situation, the data for a current patient is compared to the training data used to train a prediction model and/or to a decision function of the prediction model. The comparison indicates the training content relative to the current patient, so provides a user with information on the reliability of the prediction for the current situation. The indication deals with the variation of the data of the current patient from the training data or relative to the prediction model, allowing the user to see how well trained the predication model is relative to the current patient. This indication is in addition to any global confidence output through application of the prediction model to the data of the current patient.

Abstract: Methods, systems, and apparatus, including computer programs encoded on computer storage media, for training and deploying machine-learned classification of radio frequency (RF) signals. One of the methods includes obtaining input data corresponding to the RF spectrum; segmenting the input data into one or more samples; and for each sample of the one or more samples: obtaining information included in the sample, comparing the information to one or more labeled signal classes that are known to the machine-learning network, using results of the comparison, determining whether the information corresponds to the one or more labeled signal classes, and in response, matching, using an identification policy of a plurality of policies available to the machine-learning network, the information to a class of the one or more labeled signal classes, and providing an output that identifies an information signal corresponding to the class matching the information obtained from the sample.

Abstract: A processor-implemented data processing method includes encoding a plurality of weights of a filter of a neural network using an inverted two's complement fixed-point format; generating weight data based on values of the encoded weights corresponding to same filter positions of a plurality of filters; and performing an operation on the weight data and input activation data using a bit-serial scheme to control when to perform an activation function with respect to the weight data and input activation data.

Abstract: A method for automatically training a neural network includes at a trainer having a first communication device and a perception recorder, continuously recording the surroundings in the vicinity of the first object; receiving, at the trainer, a message from a communication device associated with an object in the vicinity of the trainer, the message including information about the position and the type of the object; identifying a recording corresponding to the time at which the message is received from the object; correlating the received positional information about the second object with a corresponding location in the recording to identify the object in the recording; classifying the identified object based on the type of information received in the message from the object; and using the classified recording to train the neural network.

Abstract: A method and system for voice emotion identification contained in audio in a call providing customer support between a customer and a service agent by implementing an emotion identification application to identify emotions captured in a voice of the customer from audio received by a media streaming device; receiving, by the emotion identification application, an audio stream of a series of voice samples contained in consecutive frames from audio received; extracting, by the emotion identification application, a set of voice emotion features from each frame in each voice sample of the audio by applying a trained machine learning (ML) model for identifying emotions utilizing a neural networks to determine one or more voice emotions by a configured set of voice emotion features captured in each voice sample; and classifying, by the emotion identification application, each emotion determined by the trained ML model based on a set of classifying features to label one or more types of emotions captured in each voic

Abstract: Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for training a student neural network. In one aspect, there is provided a method comprising: processing a training input using the student neural network to generate a student neural network output comprising a respective score for each of a plurality of classes; processing the training input using a brain emulation neural network to generate a brain emulation neural network output comprising a respective score for each of the plurality of classes; and adjusting current values of the student neural network parameters using gradients of an objective function that characterizes a similarity between: (i) the student neural network output for the training input, and (ii) the brain emulation neural network output for the training input.

Abstract: A system-on-chip (SoC) integrated circuit product includes a machine learning accelerator (MLA). It also includes other processor cores, such as general purpose processors and application-specific processors. It also includes a network-on-chip for communication between the different modules. The SoC implements a heterogeneous compute environment because the processor cores are customized for different purposes and typically will use different instruction sets. Applications may use some or all of the functionalities offered by the processor cores, and the processor cores may be programmed into different pipelines to perform different tasks.