Abstract: A method for learning parameters of a CNN using a 1×K convolution operation or a K×1 convolution operation is provided to be used for hardware optimization which satisfies KPI. The method includes steps of: a learning device (a) instructing a reshaping layer to two-dimensionally concatenate features in each group comprised of corresponding K channels of a training image or its processed feature map, to thereby generate a reshaped feature map, and instructing a subsequent convolutional layer to apply the 1×K or the K×1 convolution operation to the reshaped feature map, to thereby generate an adjusted feature map; and (b) instructing an output layer to refer to features on the adjusted feature map or its processed feature map, and instructing a loss layer to calculate losses by referring to an output from the output layer and its corresponding GT.

Abstract: A method for on-device continual learning of a neural network which analyzes input data is provided to be used for smartphones, drones, vessels, or a military purpose. The method includes steps of: a learning device, (a) sampling new data to have a preset first volume, instructing an original data generator network, which has been learned, to repeat outputting synthetic previous data corresponding to a k-dimension random vector and previous data having been used for learning the original data generator network, such that the synthetic previous data has a second volume, and generating a batch for a current-learning; and (b) instructing the neural network to generate output information corresponding to the batch. The method can be performed by generative adversarial networks (GANs), online learning, and the like. Also, the present disclosure has effects of saving resources such as storage, preventing catastrophic forgetting, and securing privacy.

Abstract: A method for updating an object detector of an autonomous vehicle to adapt the object detector to a driving circumstance is provided.

Abstract: A method for performing on-device learning of embedded machine learning network of autonomous vehicle by using multi-stage learning with adaptive hyper-parameter sets is provided. The processes include: (a) dividing the current learning into a 1-st stage learning to an n-th stage learning, assigning 1-st stage training data to n-th stage training data, generating a 1_1-st hyper-parameter set candidate to a 1_h-th hyper-parameter set candidate, training the embedded machine learning network in the 1-st stage learning, and determining a 1-st adaptive hyper-parameter set; (b) generating a k_1-st hyper-parameter set candidate to a k_h-th hyper-parameter set candidate, training the (k?1)-th stage-completed machine learning network in the k-th stage learning, and determining a k-th adaptive hyper-parameter set; and (c) generating an n-th adaptive hyper-parameter set, and executing the n-th stage learning, to thereby complete the current learning.

Abstract: A method of on-vehicle active learning for training a perception network of an autonomous vehicle is provided. The method includes steps of: an on-vehicle active learning device, (a) if a driving video and sensing information are acquired from a camera and sensors on an autonomous vehicle, inputting frames of the driving video and the sensing information into a scene code assigning module to generate scene codes including information on scenes in the frames and on driving events; and (b) at least one of selecting a part of the frames, whose object detection information satisfies a condition, as specific frames by using the scene codes and the object detection information and selecting a part of the frames, matching a training policy, as the specific frames by using the scene codes and the object detection information, and storing the specific frames and specific scene codes in a frame storing part.

Abstract: A method for explainable active learning, to be used for an object detector, by using a deep autoencoder is provided. The method includes steps of an active learning device (a) (i) inputting acquired test images into the object detector to detect objects and output bounding boxes, (ii) cropping regions, corresponding to the bounding boxes, in the test images, (iii) resizing the test images and the cropped images into a same size, and (iv) inputting the resized images into a data encoder of the deep autoencoder to output data codes, and (b) (i) confirming reference data codes corresponding to the number of the resized images less than a counter threshold by referring to a data codebook, (ii) extracting specific data codes from the data codes, (iii) selecting specific test images as rare samples, and (iv) updating the data codebook by referring to the specific data codes.

Abstract: A method of self-supervised learning for detection network using deep Q-network includes steps of: performing object detection on first unlabeled image through the detection network trained with training database to generate first object detection information and performing learning operation on a first state set corresponding to the first object detection information to generate a Q-value, if an action of the Q-value accepts the first unlabeled image, testing the detection network, retrained with the training database additionally containing a labeled image of the first unlabeled image, to generate a first accuracy, and if the action rejects the first unlabeled image, testing the detection network without retraining, to generate a second accuracy, and storing the first state set, the action, a reward of the first or the second accuracy, and a second state set of a second unlabeled image as transition vector, and training the deep Q-network by using the transition vector.

Abstract: A method for learning a recurrent neural network to check an autonomous driving safety to be used for switching a driving mode of an autonomous vehicle is provided. The method includes steps of: a learning device (a) if training images corresponding to a front and a rear cameras of the autonomous vehicle are acquired, inputting each pair of the training images into corresponding CNNs, to concatenate the training images and generate feature maps for training, (b) inputting the feature maps for training into long short-term memory models corresponding to sequences of a forward RNN, and into those corresponding to the sequences of a backward RNN, to generate updated feature maps for training and inputting feature vectors for training into an attention layer, to generate an autonomous-driving mode value for training, and (c) allowing a loss layer to calculate losses and to learn the long short-term memory models.

Abstract: A learning method for acquiring at least one personalized reward function, used for performing a Reinforcement Learning (RL) algorithm, corresponding to a personalized optimal policy for a subject driver is provided. And the method includes steps of: (a) a learning device performing a process of instructing an adjustment reward network to generate first adjustment rewards, by referring to the information on actual actions and actual circumstance vectors in driving trajectories, a process of instructing a common reward module to generate first common rewards by referring to the actual actions and the actual circumstance vectors, and a process of instructing an estimation network to generate actual prospective values by referring to the actual circumstance vectors; and (b) the learning device instructing a first loss layer to generate an adjustment reward and to perform backpropagation to learn parameters of the adjustment reward network.

Abstract: A method of adjustable continual learning of a deep neural network model by using a selective deep generative replay module is provided. The method includes steps of: a learning device (a) (i) inputting training data from a total database and a sub-database into the selective deep generative replay module to generate first and second low-dimensional distribution features, (ii) inputting binary values, random parameters, and the second low-dimensional distribution features into a data generator to generate a third training data, and (iii) inputting a first training data into a solver to generate labeled training data; (b) inputting the training data, the low-dimensional distribution features, and the binary values into a discriminator to generate a first and a second training data scores, a first and a second feature distribution scores, and a third training data score; and (c) training the discriminator, the data generator, the distribution analyzer and the solver.

Abstract: A method for learning a recurrent neural network to check an autonomous driving safety to be used for switching a driving mode of an autonomous vehicle is provided. The method includes steps of: a learning device (a) if training images corresponding to a front and a rear cameras of the autonomous vehicle are acquired, inputting each pair of the training images into corresponding CNNs, to concatenate the training images and generate feature maps for training, (b) inputting the feature maps for training into long short-term memory models corresponding to sequences of a forward RNN, and into those corresponding to the sequences of a backward RNN, to generate updated feature maps for training and inputting feature vectors for training into an attention layer, to generate an autonomous-driving mode value for training, and (c) allowing a loss layer to calculate losses and to learn the long short-term memory models.

Abstract: A method for generating a lane departure warning (LDW) alarm by referring to information on a driving situation is provided to be used for ADAS, V2X or driver safety which are required to satisfy level 4 and level 5 of autonomous vehicles. The method includes steps of: a computing device instructing a LDW system (i) to collect information on the driving situation including information on whether a specific spot corresponding to a side mirror on a side of a lane, into which the driver desires to change, belongs to a virtual viewing frustum of the driver and (ii) to generate risk information on lane change by referring to the information on the driving situation; and instructing the LDW system to generate the LDW alarm by referring to the risk information. Thus, the LDW alarm can be provided to neighboring autonomous vehicles of level 4 and level 5.

Abstract: A method for achieving better performance in autonomous driving while saving computing power, by using confidence scores representing a credibility of an object detection which is generated in parallel with an object detection process is provided.

Abstract: A learning method of a CNN (Convolutional Neural Network) for monitoring one or more blind spots of a monitoring vehicle is provided. The learning method includes steps of: a learning device, if training data corresponding to output from a detector on the monitoring vehicle is inputted, instructing a cue information extracting layer to uses class information and location information on a monitored vehicle included in the training data, thereby outputting cue information on the monitored vehicle; instructing an FC layer for monitoring the blind spots to perform neural network operations by using the cue information, thereby outputting a result of determining whether the monitored vehicle is located on one of the blind spots; and instructing a loss layer to generate loss values by referring to the result and its corresponding GT, thereby learning parameters of the FC layer for monitoring the blind spots by backpropagating the loss values.

Abstract: Processes of explainable active learning, for an object detector, by using a Bayesian dual encoder is provided. The processes include: (a) inputting test images into the object detector to generate cropped images, resizing the test images and the cropped images, and inputting the resized images into a data encoder to output data codes; (b) (b1) one of (i) inputting the test images into the object detector, applying Bayesian output embedding and resizing the activation entropy maps and the cropped activation entropy maps, and (ii) inputting resized object images and applying the Bayesian output embedding and (b2) inputting the resized activation entropy maps into a model encoder to output model codes; and (c) (i) confirming reference data codes, selecting specific test images as rare samples, and updating the data codebook, and (ii) confirming reference model codes and selecting specific test images as hard samples.

Abstract: A method for optimizing an on-device neural network model by using a Sub-kernel Searching Module is provided. The method includes steps of a learning device (a) if a Big Neural Network Model having a capacity capable of performing a targeted task by using a maximal computing power of an edge device has been trained to generate a first inference result on an input data, allowing the Sub-kernel Searching Module to identify constraint and a state vector corresponding to the training data, to generate architecture information on a specific sub-kernel suitable for performing the targeted task on the training data, (b) optimizing the Big Neural Network Model according to the architecture information to generate a specific Small Neural Network Model for generating a second inference result on the training data, and (c) training the Sub-kernel Searching Module by using the first and the second inference result.

Abstract: A method for training a deep learning network based on artificial intelligence is provided. The method includes steps of: a learning device (a) inputting unlabeled data into an active learning network to acquire sub unlabeled data and inputting the sub unlabeled data into an auto labeling network to generate new labeled data; (b) allowing a continual learning network to sample the new labeled data and existing labeled data to generate a mini-batch, and train the existing learning network using the mini-batch to acquire a trained learning network, wherein part of the mini-batch are selected by referring to specific existing losses; and (c) (i) allowing an explainable analysis network to generate insightful results on validation data and transmit the insightful results to a human engineer to transmit an analysis of the trained learning network and (ii) modifying at least one of the active learning network and the continual learning network.

Abstract: A method for performing continual learning on a classifier, in a client, capable of classifying images by using a continual learning server is provided. The method includes steps of: a continual learning server (a) inputting first hard images from a first classifier of a client into an Adversarial Autoencoder, to allow an encoder to output latent vectors from the first hard images, allow a decoder to output reconstructed images from the latent vectors, and allow a discriminator and a second classifier to output attribute and classification information to determine second hard images to be stored in a first training data set, and generating augmented images to be stored in a second training data set by adjusting the latent vectors of the reconstructed images determined not as the second hard images; (b) continual learning a third classifier corresponding to the first classifier; and (c) transmitting updated parameters to the client.

Abstract: A learning method for calculating collision probability, to be used for determining whether it is appropriate or not to switch driving modes of a vehicle capable of an autonomous driving, by analyzing a recent driving route of a driver is provided. And the method includes steps of: (a) a learning device, on condition that a status vector and a trajectory vector are acquired, performing processes of (i) instructing a status network to generate a status feature map and (ii) instructing a trajectory network to generate a trajectory feature map; (b) the learning device instructing a safety network to calculate a predicted collision probability representing a predicted probability of an accident occurrence; and (c) the learning device instructing a loss layer to generate a loss by referring to the predicted collision probability and a GT collision probability, which have been acquired beforehand, to learn at least part of parameters.

Abstract: A learning method for performing a seamless parameter switch by using a location-specific algorithm selection for an optimized autonomous driving is provided. And the method includes steps of: (a) a learning device instructing a K-th convolutional layer to apply a convolution operation to K-th training images, to thereby generate K-th feature maps; (b) the learning device instructing a K-th output layer to apply a K-th output operation to the K-th feature maps, to thereby generate K-th estimated autonomous driving source information; (c) the learning device instructing a K-th loss layer to generate a K-th loss by using the K-th estimated autonomous driving source information and its corresponding GT, and then to perform backpropagation by using the K-th loss, to thereby learn K-th parameters of the K-th CNN; and (d) the learning device storing the K-th CNN in a database after tagging K-th location information to the K-th CNN.