Samsung Sds Co., Ltd.

Yeong Dae Kwon

Jin Ho Choo

Byoung Jip Kim

Il Joo Yoon

Young June Gwon

Seong Su Kim

Yong Hyun Jeong

Seung Jai Min

Jin Yeop Chang

Jeong Hoe Goo

Du Won Park

Uk Jo

Chang Hyeon Bae

An apparatus for training a reinforcement learning model according to an embodiment includes a starting point determinator configured to determine starting points from an input value of a combinatorial optimization problem, a multi-explorer configured to generate exploration trajectories by performing exploration from each of the starting points using a reinforcement learning model, a trajectory evaluator configured to calculate an evaluation value of each of the exploration trajectories using an evaluation function of the combinatorial optimization problem, a baseline calculator configured to calculate a baseline for the input value from the evaluation value of each exploration trajectory, an advantage calculator configured to calculate an advantage of each of the exploration trajectories using the evaluation value of each exploration trajectory and the baseline, and a parameter updater configured to update parameters of the reinforcement learning model by using the exploration trajectories and the advantages of each exploration trajectory.

Apparatus and method for training reinforcement learning model for use in combinatorial optimization

A method for training an image classification model according to an embodiment includes training a feature extractor and a rotation angle classifier to predict a rotation angle of each of unlabeled first training images, training the image classification model to predict a label and rotation angle of each of labeled second training images, but predict a uniform label even though an actual rotation angle of each of the second training images is changed, generating a pseudo label based on a training image that satisfy a preset condition among unlabeled candidate images, and training the image classification model to predict a rotation angle of each of the third training images, and predict a label of each of the third training images based on the pseudo label, but predict a uniform label even though an actual rotation angle of each of the third training images is changed.

Method for training image classification model and apparatus for executing the same

An apparatus includes a modified image generator generating modified images by modifying each unlabeled image, a pre-trainer to generate a feature vector for each modified image by using an artificial neural network-based encoder and train the encoder based on the feature vector for each modified image, a pseudo-label generator to generate a feature vector for each unlabeled training image, cluster the training images based on the feature vector for each training image, and generate a pseudo-label for at least one training image among the training images based on the clustering result, and a further trainer to generate a predicted label by using the trained encoder and a classification model including a classifier to generate a predicted label for an image input to the trained encoder based on a feature vector, and train the classification model based on the pseudo-label and predicted label for the at least one training image.

Apparatus and method of image clustering

A synthetic data generation apparatus according to an embodiment includes a generator for generating synthetic data from an input value, a first discriminator learned to distinguish between actual data and the synthetic data, a second discriminator learned to distinguish between the actual data and the synthetic data while satisfying differential privacy, and a third discriminator learned to distinguish between first synthetic data which is output from the generator learned by the first discriminator and second synthetic data which is output from the generator learned by the second discriminator.

Synthetic data generation apparatus based on generative adversarial networks and learning method thereof

A method of embedding a vector representing an arrangement state of objects in a 3D space according to an embodiment includes dividing a target space into M&#xd7;N subspaces extending from a first surface among planes forming the target space having a cuboid shape, calculating, for each of the subspaces, a feature representing an object arrangement state in the subspace, inputting data representing the feature of each of the subspaces into an artificial neural network and obtaining, from the artificial neural network, an embedding vector representing an arrangement state of an object in the target space.

Growing community of inventors

Seoul, South Korea

Yeong Dae Kwon