Abstract: Techniques are provided for generating and applying a granular attention hierarchical neural network model to classify a document. In various embodiments, data indicative of the document may be obtained (102) and processed (104) into a first layer of two or more layers of a hierarchical network model using a dual granularity attention mechanism to generate first layer output data, wherein the dual granularity attention mechanism weighs some portions of the data indicative of the document more heavily. Some portions of the data indicative of the document are integrated into the hieratical network model during training of the dual granularity attention mechanism. The first layer output data may be processed (106) in the second of two or more layers of the hierarchical network model to generate second layer output data. A classification label can be generated (108) from the second layer output data.

Abstract: Technologies for image processing based on a creation workflow for creating a type of images are provided. Both multi-stage image generation as well as multi-stage image editing of an existing image are supported. To accomplish this, one system models the sequential creation stages of the creation workflow. In the backward direction, inference networks can backward transform an image into various intermediate stages. In the forward direction, generation networks can forward transform an earlier-stage image into a later-stage image based on stage-specific operations. Advantageously, this technical solution overcomes the limitations of the single-stage generation strategy with a multi-stage framework to model different types of variation at various creation stages. Resultantly, both novices and seasoned artists can use these technologies to efficiently perform complex artwork creation or editing tasks.

Abstract: The technology disclosed relates to generating ground truth training data to train a neural network-based template generator for cluster metadata determination task. In particular, it relates to accessing sequencing images, obtaining, from a base caller, a base call classifying each subpixel in the sequencing images as one of four bases (A, C, T, and G), generating a cluster map that identifies clusters as disjointed regions of contiguous subpixels which share a substantially matching base call sequence, determining cluster metadata based on the disjointed regions in the cluster map, and using the cluster metadata to generate the ground truth training data for training the neural network-based template generator for the cluster metadata determination task.

Abstract: An analysis method executed by a computer includes acquiring a refine image that maximizes a score for inferring a correct label by an inferring process using a trained model, the refine image being generated from an input image used when an incorrect label is inferred; generating a map indicating a region of pixels having the same or similar level of attention degree related to inference in the inferring process, of a plurality of pixels in the generated refine image, based on a feature amount used in the inferring process; extracting an image corresponding to a pixel region whose level in the generated map is a predetermined level, from calculated images calculated based on the input image and the refine image; and generating an output image that specifies a portion related to an inference error in the inferring process, among the calculated images, based on image processing on the extracted image.

Abstract: Methods, systems, and apparatus, including computer programs encoded on computer storage media, for training an action selection neural network that is used to control a robotic agent interacting with a real-world environment.

Abstract: An example computer-vision system to convert images includes an image converter (112) to convert a near infrared light first image (202) to form a visible light image (206), and to update a coefficient of the image converter (112) based on a difference (214), an object recognizer (102) to recognize an object (208) in the first visible light image (206), and an object recognition analyzer (210) to determine the difference (214) between the object (208) recognized in the first visible light image (206) and an object (212) associated with the near infrared light image (202).

Abstract: Systems and methods are provided for training object detectors of a neural network model with a mixture of label noise and bounding box noise. According to some embodiments, a learning framework is provided which jointly optimizes object labels, bounding box coordinates, and model parameters by performing alternating noise correction and model training. In some embodiments, to disentangle label noise and bounding box noise, a two-step noise correction method is employed. In some examples, the first step performs class-agnostic bounding box correction by minimizing classifier discrepancy and maximizing region objectness. In some examples, the second step uses dual detection heads for label correction and class-specific bounding box refinement.

Abstract: An image processing method includes: inputting a to-be-processed image into a neural network; and forming discrete feature data of the to-be-processed image via the neural network, where the neural network is trained based on guidance information, and during the training process, the neural network is taken as a student neural network; the guidance information includes: a difference between discrete feature data formed by a teacher neural network for an image sample and discrete feature data formed by the student neural network for the image sample.

Abstract: A system and a method can receive a first dataset having a first label and a first context. The system and the method can also generate, at the trained deep neural network, a second dataset having the first label and a second context according to a mapping, wherein a first mapping of the plurality of mapping comprises one or more weights of the trained deep neural network that maps data having the first label and the first context to data having a second label and the first context and a second mapping of the plurality of mapping comprises one or more weights of the trained deep neural network that maps data having a second label and the first context to data having the second label and the second context, wherein the second context is different from the first context and the second label is different from the first label.

Abstract: Various embodiments provide a method for free space optical communication performance prediction method. The method includes: in a training stage, collecting a large number of data representing FSOC performance from external data sources and through simulation in five feature categories; dividing the collected data into training datasets and testing datasets to train a prediction model based on a deep neural network (DNN); evaluating a prediction error by a loss function and adjusting weights and biases of hidden layers of the DNN to minimize the prediction error; repeating training the prediction model until the prediction error is smaller than or equal to a pre-set threshold; in an application stage, receiving parameters entered by a user for an application scenario; retrieving and preparing real-time data from the external data sources for the application scenario; and generating near real-time FSOC performance prediction results based on the trained prediction model.

Abstract: A method is provided for determining an operating state of a burner. The method includes receiving baseline characteristic data for a plurality of burner operating states. The baseline characteristic data for each burner operating state of the plurality of burner operating states comprises baseline data of a plurality of data types indicative of a corresponding burner operating state. The method also includes receiving monitoring data captured for a burner by a plurality of burner sensors. The method further includes using to machine learning to compare at least a portion of the monitoring data captured for the burner with the baseline characteristic data. The method still further includes determining an operating state of the burner based at least in part on results of comparing the at least a portion of the monitoring data with the baseline characteristic data. A corresponding apparatus and computer program product are also provided.

Abstract: A system for training a generative model and a discriminative model. The generative model generates synthetic instances from latent feature vectors by generating an intermediate representation from the latent feature vector and generating the synthetic instance from the intermediate representation. The discriminative model determines multiple discriminator scores for multiple parts of an input instance, indicating whether the part is from a synthetic instance or an actual instance. The generative model is trained by backpropagation. During the backpropagation, partial derivatives of the loss with respect to entries of the intermediate representation are updated based on a discriminator score for a part of the synthetic instance, wherein the part of the synthetic instance is generated based at least in part on the entry of the intermediate representation, and wherein the partial derivative is decreased in value if the discriminator score indicates an actual instance.

Abstract: A training device may include one or more processors configured to generate, using a data augmentation model, augmented sensor data for sensor data, the sensor data provided by a plurality of sensors, wherein the augmented sensor data comprise error states of one or more sensors of the plurality of sensors providing the sensor data, and to train an object detection model based on the augmented sensor data.

Abstract: An information management apparatus comprises a communication unit configured to communicate with a plurality of external apparatuses having learning functions, and a control unit configured to control the communication with the plurality of external apparatuses performed by the communication unit. The control unit, if supervisory data generated when a predetermined external apparatus executes a learning function is received from the predetermined external apparatus via the communication unit, selects, from among the plurality of external apparatuses, an external apparatus, other than the predetermined external apparatus, with which the supervisory data is to be shared, and performs control so that the supervisory data is transmitted to the selected external apparatus.

Abstract: The present invention provides an efficient image classification method based on structured pruning, which incorporates a spatial pruning method based on variation regularization, including steps such as image data preprocessing, inputting images to neural network, image model pruning and retraining, and new image class predication and classification. The present invention adopts a structured pruning method that removes unimportant weight parameters of the original network model and reduces unnecessary computational and memory consumptions caused by the network model in image classification to simplify the image classifier, and then uses the sparsified network model to predict and classify new images. The simplified method according to the present invention improves the original network model in image classification efficiency by nearly two times, costs about 30% less memory consumption and produces a better classification result.

Abstract: A system includes a camera configured to obtain image information from objects. The system also includes a processor in communication with the camera and programmed to receive an input data including the image information, encode the input via an encoder, obtain a latent variable defining an attribute of the input data, generate a sequential reconstruction of the input data utilizing at least the latent variable and an adversarial noise, obtain a residual between the input data and the sequential reconstruction utilizing a comparison of at least the input and the reconstruction to learn a mean shift in latent space, and output a mean shift indicating a test result of the input compared to the adversarial noise based on the comparison.

Abstract: A data generation device includes one or more processors. The processors input input data into a neural network and obtain an inference result of the neural network The processors calculate a first loss and a second loss. The first loss becomes smaller in value as a degree of matching between the inference result and a target label becomes larger. The target label indicates a correct answer of the inference. The second loss is a loss based on a contribution degree to the inference result of a plurality of elements included in the input data and the target label. The processors update the input data based on the first loss and the second loss.

Abstract: An embodiment of the present disclosure provides an artificial intelligence apparatus for generating training data including a memory configured to store an artificial intelligence model, an input interface including a microphone or a camera, and a processor configured to receive, via the input interface, input data, generate an inference result corresponding to the input data by using the artificial intelligence model, receive feedback corresponding to the inference result, determine suitability of the input data and the feedback for updating the artificial intelligence model, and generate training data based on the input data and the feedback if the input data and the feedback are determined as data suitable for updating of the artificial intelligence model.

Abstract: An artificial intelligence neural network apparatus, comprising: a labeled learning database having data of a feature vector composed of N elements; a first feature vector image converter configured to visualize the data in the learning database to form an imaged learning feature vector image database; a deep-learned artificial intelligence neural network configured to use a learning feature vector image in the learning feature vector image database to perform an image classification operation; an inputter configured to receive a test image, and generate test data based on the feature vector; and a second feature vector image converter configured to visualize the test data and convert the visualized test data into a test feature vector image. The deep-learned artificial intelligence neural network is configured to determine a class of the test feature vector image.

Abstract: A calculation method using pixel-channel shuffle convolutional neural network is provided. In the method, an operating system receives original input data. The original input data is pre-processed by a pixel shuffle process to be separated into multiple groups in order to minimize dimension of the data. The multiple groups of data are then processed by a channel shuffle process so as to form multiple groups of new input data selected for convolution operation. The unselected data are abandoned. Therefore, the dimension of the input data can be much effectively minimized. A multiplier-accumulator of the operating system is used to execute convolution operation using a convolution kernel and the multiple new groups of input data. Multiple output data are then produced.