Abstract: A computer-implemented method for frequency coding of image data from an imaging sensor. The method includes: supplying first image data of an individual image recorded by an imaging sensor, the first image data having depth values of the individual image coded as a whole number or as a floating-point number; receiving the first image data by an algorithm, which frequency codes the depth values of the individual image by a predefined number of periodic functions; and outputting second image data by the algorithm, the second image data having frequency coded depth values of the individual image. A computer-implemented method is described for supplying an algorithm of machine learning for the classification of objects included in image data of an individual image from an imaging sensor. A system for the frequency coding of image data from an imaging sensor, a computer program, and a computer-readable data carrier, are also described.

Abstract: A microscopy system for generating training data for a machine learning model comprises a microscope configured to capture an image. The microscopy system further comprises a computing device configured to generate a segmentation mask based on the image, adjust a pattern described by a parameterized model to the segmentation mask, generate an updated segmentation mask using the adjusted pattern, and incorporate the updated segmentation mask or an image derived from the same in the training data.

Abstract: In part, the disclosure relates to methods, and systems suitable for evaluating image data from a patient on a real time or substantially real time basis using machine learning (ML) methods and systems. Systems and methods for improving diagnostic tools for end users such as cardiologists and imaging specialists using machine learning techniques applied to specific problems associated with intravascular images that have polar representations. Further, given the use of rotating probes to obtain image data for OCT, IVUS, and other imaging data, dealing with the two coordinate systems associated therewith creates challenges. The present disclosure addresses these and numerous other challenges relating to solving the problem of quickly imaging and diagnosis a patient such that stenting and other procedures may be applied during a single session in the cath lab.

Abstract: A computer-implemented method of automatically labeling medical images is provided. The method includes clustering training images and training labels into clusters, each cluster including a representative template having a representative image and a representative label. The method also includes training a neural network model with a training dataset that includes the training images and the training labels, and target outputs of the neural network model are labels of the medical images. The method further includes generating a suboptimal label corresponding to an unlabeled test image using the trained neural network model, and generating an optimal label corresponding to the unlabeled test image using the suboptimal label and representative templates.

Abstract: Event data and event participant data corresponding to a time during an event are encoded into a multidimensional feature vector using a trained encoder network. Using an attention mask, the multidimensional feature vector is adjusted, the adjusting amplifying a portion of the multidimensional feature vector according to an importance level of the portion. The adjusted multidimensional feature vector is decoded into an excitement level score using a trained decoder network. Using the excitement level score and a trained neural network model, a frequency and an amplitude of simulated crowd noise corresponding to the time during the event are generated.

Abstract: A training data generation apparatus (10) according to the present invention includes a noise determination unit (11) that determines whether or not training data that is to be used in machine learning includes noise, and a noise addition unit (12) that generates new training data by adding noise to training data that has been determined by the noise determination unit (11) as not including noise.

Abstract: An automatic generation system of a training image and a method thereof are provided. The disclosure generates a training image and records the target category and the target position. The disclosure adds the target image to the container image as a candidate image, calculates a reliability of the candidate image, and repeatedly executes the process until the reliability of the candidate image meets a threshold condition for generating the training image. The disclosure is able to generate the training images automatically, and the recognition difficulty of the training image is adjustable by the user, so as to be suitable for customized recognition training.

Abstract: A first learning unit performs learning of a first neural network that extracts a feature vector in each pixel of a target image including a plurality of objects and that outputs a feature map in which feature vectors of pixels belonging to individual objects included in the target image are clustered and distributed as a plurality of the feature vector groups in a feature space which is a space of the feature vector. A second learning unit performs learning of a second neural network that outputs a class classification result of a plurality of objects belonging to the same category included in the target image in response to input of the feature vector of the target image.

Abstract: Machine learning based models recognize objects in images. Specific features of the object are extracted from the image using machine learning based models. The specific features extracted from the image assist deep learning based models in identifying subtypes of a type of object. The system recognizes the objects and collections of objects and determines whether the arrangement of objects violates any predetermined policies. For example, a policy may specify relative positions of different types of objects, height above ground at which certain types of objects are placed, or an expected number of certain types of objects in a collection.

Abstract: Techniques for training a machine learned (ML) model to determine depth data based on image data are discussed herein. Training can use stereo image data and depth data (e.g., lidar data). A first (e.g., left) image can be input to a ML model, which can output predicted disparity and/or depth data. The predicted disparity data can be used with second image data (e.g., a right image) to reconstruct the first image. Differences between the first and reconstructed images can be used to determine a loss. Losses may include pixel, smoothing, structural similarity, and/or consistency losses. Further, differences between the depth data and the predicted depth data and/or differences between the predicted disparity data and the predicted depth data can be determined, and the ML model can be trained based on the various losses. Thus, the techniques can use self-supervised training and supervised training to train a ML model.

Abstract: Method, apparatuses, systems, electronic devices, computer readable storage media, and computer program products for controlling image acquisition are provided. In one aspect, a method includes: providing a first image sample set to a first neural network; selecting one or more first hard samples from the first image sample set according to a processing result of the first neural network for each first image sample in the first image sample set; determining acquisition environment information of the one or more first hard samples based on the one or more first hard samples; and generating, according to the acquisition environment information, image acquisition control information for instruction of an acquisition of a second image sample set comprising one or more second hard samples.

Abstract: An apparatus for evaluating the performance of a deep learning model according to an embodiment may include an image processor configured to generate N (N?2) different second image data through data augmentation of first image data that is not labeled and transmit the generated second image data to a deep learning model, and an analyzer configured to analyze whether the deep learning model has output a correct answer by receiving N output data obtained by predicting each of the N second image data into a specific class from the deep learning model.

Abstract: A microscopy system comprises a microscope which is set up to record at least one microscope image and a computing device which comprises a trained image processing model set up to calculate an image processing result on the basis of the at least one microscope image. The computing device is set up to verify the trained image processing model by: receiving a validation image and an associated target image; entering the validation image into the trained image processing model, which calculates an output image therefrom; entering image data based on at least the output image and the associated target image into a trained verification model which is trained to calculate an evaluation that indicates a quality that depends on the image data for entered image data; and calculating an evaluation by the trained verification model on the basis of the entered image data. A method for verifying a trained image processing model is additionally described.

Abstract: Systems and methods for diagnosing and testing fairness of machine learning models based on detecting individual violations of group definitions of fairness, via adversarial attacks that aim to perturb model inputs to generate individual violations. The systems and methods employ auxiliary machine learning models using a local surrogate for identifying group membership and assess fairness by measuring the transferability of attacks from this model. The systems and methods generate fairness indicator values indicative of discrimination risk due to the target predictions generated by the machine learning model, by comparing gradients of the machine learning model to gradients of an auxiliary machine learning model.

Abstract: Methods, systems, and apparatus, including computer programs encoded on computer storage media, for training a machine learning model. One of the methods includes obtaining a training data set for training a machine learning model, the training data set comprising a plurality of training inputs; determining a plurality of data augmentation policies, wherein each data augmentation policy defines a procedure for processing a training input to generate a transformed training input; for each data augmentation policy, training the machine learning model using the data augmentation policy; determining, for each data augmentation policy, a quality measure of the machine learning model that has been trained using the data augmentation policy; and selecting a final data augmentation policy based using the quality measures of the machine learning models.

Abstract: A computer-implemented method and computer program product for predicting an impact of an adjustment to a machine learning model to key performance indicators, and a forecasting engine. The computer-implemented method may comprise receiving a proposed adjustment to a machine learning model, calculating, using a regression machine learning model to ingest the proposed adjustment, a set of value components for a key performance indicator (KPI), calculating a plurality of results for the KPI using the set of value components, automatically determining whether the plurality of results exceeds a performance threshold, and recommending the proposed adjustment based on the determination.

Abstract: Method and server for determining a target combination of metric-specific thresholds to be used with a plurality of nested metrics for performing binary classification of a digital object are disclosed. The method includes acquiring object-specific validation datasets, and a plurality of nested metrics thereon, thereby generating a plurality of prediction values. During a first iteration, the server compares predictions values against a first combination of metric-specific thresholds and generates first precision parameters and first recall parameters for the first iteration. During a second iteration, the server adjusts one of the first combination thereby generating a second combination, compares the predictions values against the second combination, and generates second precision parameters and second recall parameters for the second iteration. The method includes selecting, by the one of the first combination and the second combination as the target combination of metric-specific thresholds.

Abstract: The fine-grained variations in product images are usually due to slight variations in text, size, and color of the package. Both marginal variations in image content and illumination poses an important challenge in product classification. This disclosure relates to a system and method for fine-grained classification of similar-looking products utilizing object-level and part-level information. The system simultaneously captures an object-level and part-level information of the product. The object-level classification score of the product is estimated with the trained RC-Net, a deep supervised convolutional autoencoder. For annotation-free modelling of part-level information of the product the discriminative part-proposal of the product is identified around the BRISK key points. An ordered sequence of the discriminative part-proposals and the product image, encoded using stacked convolutional LSTM network, estimates the part-level classification score.

Abstract: In variants, the method can include: sampling cavity measurements of the cook cavity; determining an image representation using the cavity measurements; determining a food class based on the image representation; optionally comparing the image representation to prior image representations; optionally determining a new food class based on the image representation; optionally updating a food identification module with the new food class; and optionally determining a cook program associated with the new food class.

Abstract: A method of modelling data, comprising: training an objective function of a linear classifier, based on a set of labeled data, to derive a set of classifier weights; defining a posterior probability distribution on the set of classifier weights of the linear classifier; approximating a marginalized loss function for an autoencoder as a Bregman divergence, based on the posterior probability distribution on the set of classifier weights learned from the linear classifier; and classifying unlabeled data using the autoencoder according to the marginalized loss function.

Abstract: A system receives an observation probability distribution function associated with a target object that was detected by sensors of an autonomous vehicle. The system identifies a target attribute of the target object, and detects a target attribute value associated with the target object. The system determines a first probability distribution function representing a probability of the autonomous vehicle detecting an object having an object label, determines a second probability distribution function defining a probability of the autonomous vehicle detecting the target attribute, determines a third probability distribution function defining a probability of the target attribute being present for the target object based on the target attribute value, and determines an attribute probability distribution function defining a probability that the target attribute is actually present for the target object.

Abstract: Embodiments may include novel techniques for intermediate model generation using historical data and domain knowledge for Reinforcement Learning (RL) training. Embodiments may start with gathering client data.

Abstract: An object is to match the texture of a printed material, in addition to the tint, between different printing devices. In order to implement this object, a printed material output by a destination device is scanned and whether a certain area in image data obtained by the scan is an area printed in K color on the printed material or an area printed in CMY color mixture is determined. Then, based on the determination results, a printing parameter is determined.

Abstract: In a printing device, a supply portion is configured to convey a tape in its longitudinal direction. The tape includes: a plurality of labels arranged continuously in the longitudinal direction; and a plurality of storage elements provided on respective ones of the plurality of labels. A first storage element is provided on a first label and configured to store first authentication data. The second storage element is provided on a second label and configured to store second authentication data. A printing portion is configured to print on the plurality of labels. A controller is configured to perform: reading the first authentication data from the first storage element and the second authentication data from the second storage element by a reading portion; and determining whether the first authentication data is correlated to the second authentication data to meet an authentication condition.

Abstract: A durable pet tag having machine readable indicia. A multiple layer label is pre-assembled and applied and embedded into a metal substrate for durability. A two sided pet tag having machine readable indicia on a first side and human readable indicia on a second side is provided, the human readable indicia being stamped into the side of the substrate opposite the machine readable printed sheet so that a single pet tag is provided with machine readable data to allow finders of lost pets to retrieve owner information using a smart phone and the database accessible with the pet tag machine readable indicia. The pet tag also has human readable indicia to comply with legal requirements.

Abstract: The invention relates to a chip card designed to communicate data in a contactless mode with a card reader operating at a reading frequency. The resonance frequency of the chip card may change according to the capacitance of the chip used in the contactless mode of the chip card. In order to be able to use various chips without changing the booster antenna design, the card antenna circuit is provided with a capacitance element such that the chip card including the card antenna circuit and the chip module has two different resonance frequencies, one of which being equal to, or lower than, the reading frequency and the other being equal to, or greater, than the reading frequency. This create a broadband wherein the reading frequency falls.

Abstract: A method for generating an Artificial Intelligence (AI) bot comprises the steps of receiving (210) information of a human (3102), indicative of the physical characteristics including an appearance and vocals of the human (3102) and behavioural characteristics of the human (3102), analysing (220) the information for identifying and mimicking the vocals of the human (3102), analysing (230) the information for identifying and imitating the appearance of the human (3102), generating (240) the AI bot having the appearance of the human (3102) in a mixed reality space, processing and merging (250) the identified physical characteristics and the behavioural characteristics into the AI bot, displaying (260) the AI bot having physical characteristics and the behavioural characteristics of the human (3102) using the HMD (102), enabling (270) an interaction of the AI bot with users in the mixed reality space, thereby enabling the omnipresence of the human (3102).

Abstract: An example fused convolutional layer, comprising, a comparator capable of reception of a first zero point and a multiply-accumulation result, a first multiplexer coupled to the comparator, wherein the first multiplexer receives a plurality of power-of-two exponent values, a shift normalizer, coupled to the first multiplexer, wherein the shift normalizer is capable of receiving the multiply-accumulation result and the plurality of power-of-two exponent values, wherein the shift normalizer limits a quantization of the multiply-accumulation result to a power-of-two scale and a second multiplexer coupled to an output of the shift normalizer, the first multiplexer and receives a second zero point and outputs an activation.

Abstract: A method of unification based coding for neural network model compression is performed by at least one processor and includes receiving a layer uniform flag indicating whether a quantized weight of an input neural network is encoded using a uniform coding method, and determining whether the quantized weight is encoded using the uniform coding method, based on the received layer uniform flag. The method further includes, based on the quantized weight being determined to be encoded using the uniform coding method, encoding the quantized weight, using the uniform coding method, and based on the quantized weight being determined to not be encoded using the uniform coding method, encoding the quantized weight, using a non-uniform coding method.

Abstract: Aspects of the present disclosure describe improved artificial neural network architectures for resource constrained application that employ tiny skips or improved parameter efficiency of existing artificial neural network architectures designed for resource-constrained applications by employing content-based interaction layers. Our technique is demonstrated with a specific example in which we replace spatial convolution layers in a MobilenetV2-like structure with Lambda Layers and achieve a significant improvement in accuracy while using the same number of parameters.

Abstract: A method for determining safety-critical output values by way of a data analysis device for a technical entity. The method includes receiving data and/or measured values for the entity by way of the data analysis device. The method further includes processing the input values (x) by way of the data analysis device using a software application in order to determine at least one first output value (y1). The method further includes using a neural network having a plurality of layers (h?(x)) with first learnable parameters (?); modifying a layer of the neural network using a function (?) so that the output value (y1) is located within a defined value range (C(s)) of at least one target parameter (s); and determining the target parameter (s) and/or the value range (C(s)) of the target parameter (s) using further additional layers (k?(x)) of the neural network with second learnable parameters (?).

Abstract: An exemplary embodiment provides an explanation and interpretation generation system for creating explanations in different human and machine-readable formats from an explainable and/or interpretable machine learning model. An extensible explanation architecture may allow for seamless third-party integration. Explanation scaffolding may be implemented for generating domain specific explanations, while interpretation scaffolding may facilitate the generation of domain and scenario specific interpretations. An exemplary explanation filter interpretation model may provide an explanation and interpretation generation system optional filtering and interpretation filtering and briefing capabilities. An embodiment may cluster explanations into concepts to incorporate information such as taxonomies, ontologies, causal models, statistical hypotheses, data quality controls, domain specific knowledge and allow for collaborative human knowledge injection.

Abstract: A machine learning device performing online learning of input data of one or more dimensions aligned in a pre-determined order using a recurrent neural network having a plurality of nodes connected by edges to which weights are assigned performs an output data generating process and a weight updating process every time the input layer receives the input data of one or more dimensions in the pre-determined order, in which, the weight updating process is a process in which a weight assigned to each edge connecting a 1st intermediate node and a 2nd intermediate node and a weight assigned to each edge connecting a 2nd intermediate node and an output node are updated using an equation derived based on an extended Kalman filter method, 1st intermediate data of one or more dimensions, and output data of one or more dimensions.

Abstract: A classification device and a classification method based on a neural network are provided. A heterogeneous integration module includes a convolutional layer, a data normalization layer, a connected layer and a classification layer. The convolutional layer generates a first feature map according to a first image data. The data normalization layer normalizes a first numerical data to generate a first normalized numerical data. The first numerical data corresponds to the first image data. The connected layer generates a first feature vector according to the first feature map and the first normalized numerical data. The classification layer generates a first classification result corresponding to a first time point according to the first feature vector. The heterogeneous integration module generates a second classification result corresponding to a second time point.

Abstract: A processor-implemented neural network operation method includes: receiving a feature map on which a neural network operation is to be performed; selecting a predetermined area from the feature map; generating a normalization parameter based on the predetermined area; and performing the neural network operation based on the normalization parameter.

Abstract: The present invention relates to a method of estimating lithium battery capacity based on a convolution long-short-term memory neural network (CNN-LSTM). The present invention obtains a model that lithium battery capacity estimation through the four steps: processing a lithium battery's data, selecting parameters of an improved convolution long-short-term memory neural network using a genetic algorithm, training the improved CNN-LSTM, and testing model. Hyper-parameters of the improved CNN-LSTM are optimized using the genetic algorithm. Using the convolution neural network to extract the spatial features of lithium battery charge and discharge data, and then input these features into the improved long-short-term memory neural network to extract temporal features, estimated capacity is output through a fully connected layer finally. The present invention overcomes the limitation of the traditional model-based algorithm overly relying on the battery model and has the engineering application prospect.

Abstract: A computer-implemented method for a neural network, for example an artificial deep neural network. The method includes: providing a plurality of training data sets, each training data set comprising input data for the neural network and associated output data, training the neural network based on the plurality of training data sets and a loss function, wherein the loss function is based on a bit-wise correlation of an output value provided by the neural network and a predetermined function characterizing an operation of a physical system.

Abstract: To configure a set of user devices, which comprises one or more user devices, to monitor a subset of beams, at least one past beam sequence indicating one or more beams, which served the set is determined, and inputted to a trained model which outputs a probability distribution. Then as many beams as is a number of beams determined for the set to monitor is selected from the probability distribution according to a first criteria, and the set of user devices is configured to monitor and report beams in the beam group. Past beam sequences are also used in training. From the past sequences, set-specifically, past beams that served a set within a first time interval are determined to be used as training data, and future beams that served the set within a second time interval following the first time interval are determined to be used as validation data.

Abstract: Methods, processing units and media for multi-bandwidth separated feature extraction convolution in a neural network are described. A convolution block splits input channels of an activation map into multiple branches, each branch undergoing convolution at a different bandwidth by using down-sampling of the inputs. The outputs are concatenated by up-sampling the outputs of the low-bandwidth branches using pixel shuffling. The concatenation operation may be a shuffled concatenation operation that preserves separated multi-bandwidth feature information for use by subsequent layers of the neural network. Embodiments are described which apply frequency-based and magnitude-based attention to the weights of the convolution kernels based on the frequency band locations of the weights.

Abstract: A system and method for performing sets of multiplications in a manner that accommodates outlier values. In some embodiments the method includes: forming a first set of products, each product of the first set of products being a product of a first activation value and a respective weight of a first plurality of weights. The forming of the first set of products may include multiplying, in a first multiplier, the first activation value and a least significant sub-word of a first weight to form a first partial product; multiplying, in a second multiplier, the first activation value and a least significant sub-word of a second weight; multiplying, in a third multiplier, the first activation value and a most significant sub-word of the first weight to form a second partial product; and adding the first partial product and the second partial product.

Abstract: A neural network operation apparatus includes: a receiver configured to receive a first input feature map; a controller configured to control multiplier-accumulators (MACs) included in a first MAC array; and a first operation engine comprising the first MAC array and configured to process the first input feature map based on the MACs of which operation states are controlled.

Abstract: A neural network apparatus includes: a plurality of memory cells each comprising a variable resistance element and a first transistor; a plurality of bit lines extending in a first direction; and a plurality of word lines extending in a second direction, crossing the bit lines and respectively connected to the first transistor of the plurality of memory cells; a plurality of sub-column circuits each comprising memory cells of the memory cells connected in parallel along the first direction; and a column circuit comprising two or more of the sub-column circuits connected in series along the second direction, wherein, when a neural network operation is performed, the column circuit outputs a summation current to a bit line connected to the column circuit based on voltage applied to the plurality of word lines.

Abstract: Disclosed is an operation method of a neural network element using a Hall voltage. The neural network element has a hole pattern portion, and the hole pattern portion has a cross shape. When a pulse current is applied, horizontal magnetic anisotropy is formed in a ferromagnetic layer by means of spin-orbit torque, and when an external magnetic field in a direction perpendicular to the pulse current is applied, the inversion of magnetization occurs by means of additional torque. The movement of a magnetic domain wall is performed by the inversion of magnetization, spin electrons applied thereby are scattered, and a Hall voltage is generated according to the anomalous Hall effect. The generated Hall voltage increases according to the number of applications of the pulse current or pulse voltage.

Abstract: The present disclosure relates to a method and a device for generation operation data and a related product, the product comprising a controller unit, and the controller unit comprising: an instruction caching unit, an instruction processing unit and a queue-storing unit. The instruction caching unit is used to store computing instructions associated with artificial neural network operations. The instruction processing unit is used to resolve the computing instructions to obtain a plurality of operation instructions. The queue-storing unit is used to store an instruction queue, which comprises: a plurality of operation instructions or computing instructions to be executed according to the front-to-rear sequence of the queue. Through the above method, the present disclosure may improve the operation efficiency of the related product when carrying out neural network model operations.

Abstract: One embodiment provides an apparatus comprising an instruction cache to store a plurality of instructions, a scheduler unit coupled to the instruction cache, the scheduler unit to schedule the plurality of instructions for execution, an instruction fetch and decode unit to decode the plurality of instructions to determine a set of operations to perform in response, one or more compute blocks to perform parallel multiply-accumulate operations based on the instruction fetch and decode unit decoding a first instruction of the plurality of instructions, and matrix multiplication logic to perform matrix multiplication operations based on the instruction fetch and decode unit decoding a second instruction of the plurality of instructions.

Abstract: An integrated-circuit neural network includes chain of multiply-accumulate units co-located with a high-bandwidth storage array. Each multiply accumulate includes a digital input port, analog input port and multiply-adder circuitry. The digital input port receives a matrix of digital-weight values from the storage array and the analog input port receives a counterpart matrix of analog input signals, each analog input signal exhibiting a respective electronic current representative of input value. The multiply-adder circuitry generates a matrix of analog output signals by convolving the matrix of digital-weight values with the matrix of analog input signals including, for each analog output signal within the matrix of analog output signals, switchably enabling weighted current contributions to the analog output signal based on logic states of on respective bits of one or more of the digital-weight values.

Abstract: Embodiments may relate to a structure to be used in a neural network. A first column and a second column, both of which are to couple with a substrate. A capacitor structure may be electrically coupled with the first column. An insulator-metal transition (IMT) structure may be coupled with the first column such that the capacitor structure is electrically positioned between the IMT structure and the first column. A resistor structure may further be electrically coupled with the IMT structure and the second column such that the resistor structure is electrically positioned between the second column and the IMT structure. Other embodiments may be described or claimed.

Abstract: A vehicle includes one or more sensors configured to obtain raw data related to a scene, one or more processors, and machine readable instructions stored in one or more memory modules. The one machine readable instructions, when executed by the one or more processors, cause the vehicle to: process the raw data with a first neural network stored in the one or more memory modules to obtain a first prediction about the scene, transmit the raw data to a computing device external to the vehicle, receive a second prediction about the scene from the computing device in response to transmitting the raw data to the computing device, and determine an updated prediction about the scene based on a combination of the first prediction and the second prediction.

Abstract: Performing a goal-seek analysis of spatial-temporal data by generating a hierarchical cluster according to spatial temporal data, determining a spatial-temporal location input for a target, determining spatial-temporal predictor values for the spatial-temporal location, and adjusting the hierarchical cluster according to and the spatial-temporal predictors.

Abstract: Disclosed are systems and methods to incrementally train neural networks. Incrementally training the neural networks can include defining a probability distribution of labeled training examples from a training sample pool, generating a first training set based off the probability distribution, training the neural network with the first training set, adding at least one additional training sample to the training sample pool, generating a second training set, and training the neural network with the second training set. The incremental training can be recursive for additional training sets until a decision to end the recursion is made.