Abstract: A method for controlling a loss scale by class includes identifying class-wise losses of a neural network, analyzing a relative correlation by normalizing the identified class-wise losses, estimating class-wise loss scaling weights according to a result of the analyzing of the relative correlation, and calculating a weighted sum of the class-wise losses by applying the estimated class-wise loss scaling weights to the identified class-wise losses of the neural network, and deriving a total loss of the neural network by summing the identified class-wise losses and the weighted sum of the class-wise losses.

Abstract: According to an embodiment of the present disclosure, there is provided a waste plastic gasification device including: a raw material input unit for inputting waste plastic raw materials; a first gasification unit that pyrolyzes the waste plastic raw materials at low temperatures using superheated steam to produce a mixed gas and preheats residual raw materials that have not been decomposed; a second gasification unit that further decomposes the preheated residual raw materials through plasma to produce the mixed gas; a residue discharge unit that stores and discharges waste plastic residual melt slag that has not been decomposed in the second gasification unit; and a heat exchange unit for cooling the mixed gas generated in the first gasification unit and the second gasification unit, wherein the second gasification unit is inclined in a direction of the residue discharge unit to discharge the waste plastic residual melt slag.

Abstract: An embodiment provides an alkali metal thermal-to-electric converter including a thermal-to-electric conversion cell including three layers including an anode layer, a solid electrolyte layer, and a cathode layer, having a convex-concave shape with alternately appearing concave and convex portions, and configured to move alkali metal ions through the solid electrolyte layer, a high temperature portion that supplies a high temperature alkali metal fluid to the anode layer of the thermal-to-electric conversion cell, and a low temperature portion that condenses the alkaline metal fluid discharged to the cathode layer of the thermal-to-electric conversion cell to a low temperature and moves the alkaline metal fluid to the high temperature portion.

Abstract: An apparatus for manufacturing hydrogen peroxide water by using electrostatic spraying is disclosed. An apparatus for manufacturing hydrogen peroxide water by using electrostatic spraying according to a first embodiment of the present disclosure comprises: a nozzle unit for spraying reaction water; a ground unit disposed opposite to the nozzle unit; and a power supply unit for applying a high voltage between the nozzle unit and the ground unit, wherein the reaction water is formed into fine droplets with a particle size of 20 ?m or less in the nozzle unit and is electrostatically sprayed, and ionized hydrogen ions (H+) and hydroxide ions (OH?) react with each other again to form hydrogen peroxide.

Abstract: Disclosed herein are an object recognition apparatus and method based on environment matching. The object recognition apparatus includes memory for storing at least one program, and a processor for executing the program, wherein the program performs extracting at least one key frame from a video that is input in real time, determining a similarity between the key frame extracted from the input video and each of videos used as training data of prestored multiple recognition models, based on a pretrained similarity-matching network, and selecting a recognition model pretrained with a video having a maximal similarity to the key frame extracted from the input video, preprocessing the input video such that at least one of color and size of a video used as training data of an initial model is similar to that of the input video, and recognizing the preprocessed video based on the initial model.

Abstract: Disclosed herein is a method for detecting an anomaly state based on screen output. The method includes receiving the output screen of a target device to be monitored, setting a target region to be examined in the output screen of the target device to be monitored, calculating a feature value vector corresponding to the state of the target region to be examined, calculating an anomaly score using a pretrained auto-encoder by receiving the feature value vector as input, and determining whether the target device to be monitored is anomalous using the anomaly score.

Abstract: An apparatus for classifying fake images according to an embodiment of the present disclosure includes an artifact remover configured to receive an input image to generate an artifact-removed image from which artifacts are removed, an artifact image generator configured to generate an artifact image by using a difference between the input image and the artifact removal image, and a determiner configured to determine whether the artifact image is a real image or a fake image by receiving the artifact image.

Abstract: An apparatus for video representation learning according to an embodiment may extract video features from video data to generate a video embedding, extract image features from image data extracted from the video data to generate an image embedding, and extract audio features from audio data extracted from the video data to generate an audio embedding. Further, contrastive learning may be performed by generating a first compositional embedding based on the video embedding and the audio embedding, generating a second compositional embedding based on the video embedding and the audio embedding, generating a positive sample and a negative sample based on a correlation between the image embedding and the audio embedding, and then using the data.

Abstract: An apparatus for video representation learning according to an embodiment may extract video features from video data to generate a video embedding, extract image features from image data extracted from the video data to generate an image embedding, and extract audio features from audio data extracted from the video data to generate an audio embedding. Further, contrastive learning may be performed by generating a first compositional embedding based on the video embedding and the audio embedding, generating a second compositional embedding based on the video embedding and the audio embedding, generating a positive sample and a negative sample based on a correlation between the image embedding and the audio embedding, and then using the data.

Abstract: A method and apparatus for training a fake image discriminative model according to an embodiment of the present disclosure includes generating one or more fake images for a real image by selecting one or more encoding layers and one or more decoding layers from a generator network of an autoencoder structure, generating a training image set based on the one or more fake images, and training a classifier for discriminating a fake image by using the training image set.

Abstract: A method for training a deep neural network according to an embodiment includes training a deep neural network model using a first data set including a plurality of labeled data and a second data set including a plurality of unlabeled data, assigning a ground-truth label value to some of the plurality of unlabeled data, updating the first data set and the second data set such that the data to which the ground-truth label value is assigned is included in the first data set, and further training the deep neural network model using the updated first data set and the updated second data set.

Abstract: Disclosed herein are an object recognition apparatus and method based on environment matching. The object recognition apparatus includes memory for storing at least one program, and a processor for executing the program, wherein the program performs extracting at least one key frame from a video that is input in real time, determining a similarity between the key frame extracted from the input video and each of videos used as training data of prestored multiple recognition models, based on a pretrained similarity-matching network, and selecting a recognition model pretrained with a video having a maximal similarity to the key frame extracted from the input video, preprocessing the input video such that at least one of color and size of a video used as training data of an initial model is similar to that of the input video, and recognizing the preprocessed video based on the initial model.

Abstract: A machine learning apparatus according to an embodiment includes a feature extractor configured to extract features from an object region of an image, a label processor configured to create sentence label embeddings from a sentence label corresponding to the object region, a first training data creator to extract first sub-features from a plurality of first sub-regions created by partitioning the object region, add the sentence label embeddings to the extracted first sub-features, and add the first sub-features added with the sentence label embeddings to the features of the object region, a second training data creator to extract a plurality of second sub-regions along a bounding surface of the object region, create an attention matrix from the second sub-regions, and create a training data by applying the attention matrix to the features of the object region, and a trainer to train an object detection model using the training data.

Abstract: A method for predicting a value for unlabeled data according to an embodiment is executed in a computing device including one or more processors and a memory storing one or more programs executed by the one or more processors. The method includes training a relation estimation model for predicting a label for labeled data based on a relation between respective predicted values of a plurality of pre-trained predictive models for the labeled data and a target model and obtaining a predicted value for unlabeled data based on the relation estimation model.

Abstract: An apparatus for deep fake image discrimination according to an embodiment includes an interface unit configured to receive image data and a classifier configured to determine whether the image data input through the interface unit is a deep fake image, and the classifier is trained to determine a deep fake image based on a synthetic image generated by swapping a portion of a real image with a fake image generated by self-replicating the real image.

Abstract: A compound of Chemical Formula 1, and an organic photoelectric device, an image sensor, and an electronic device including the same are disclosed: In Chemical Formula 1, each substituent is the same as described in the detailed description.

Abstract: Provided are an optimization system and method of a deep learning model. According to example embodiments, by optimizing the structure of the deep learning model appropriately for a target dataset without fixing the structure of the deep learning model, it is possible to generate a model structure capable of having high performance on the target dataset and also saving resources.

Abstract: An apparatus for classifying fake images according to an embodiment of the present disclosure includes an artifact remover configured to receive an input image to generate an artifact-removed image from which artifacts are removed, an artifact image generator configured to generate an artifact image by using a difference between the input image and the artifact removal image, and a determiner configured to determine whether the artifact image is a real image or a fake image by receiving the artifact image.

Abstract: An image processing apparatus includes a superpixel extractor configured to extract a plurality of superpixels from an input original image, a backbone network including N feature extracting layers (here, N is a natural number of two or more) which divide the input original image into grids including a plurality of regions and generate an output value including a feature value for each of the divided regions, and a superpixel pooling layer configured to generate a superpixel feature value corresponding to each of the plurality of superpixels using a first output value to an Nth output value output from each of the N feature extracting layers.

Abstract: A thrust supporting apparatus includes: a thrust runner configured to rotate integrally with a shaft; and a foil thrust bearing supporting an axial load of the shaft, the axial load being transmitted from the thrust runner. The foil thrust bearing includes a top foil facing a first surface of the thrust runner, and the thrust runner comprises first one or more grooves formed on the first surface.