We present a new multi-domain adversarial transformer for ISO domain transformation to improve the quality of low-light photographs.

We propose a new approach for pixel-wise image enhancement non-locally and a new metric for examining RGB color image quality.

In this thesis, I applied a Variational Autoencoder (VAE), a deep generative model, to automatically generate more realistic and diverse dances from music.

We propose two methods for controllable generation of complex graphic layouts. The first method uses a pre-trained deep generative model to generate layouts that satisfy given constraints. The second method represents visual containment as a tree structure and generates web page layouts. Finally, we summarize design guidelines for controllable generation methods and discuss future directions.

We propose to generate brand new Super Mario Bros(SMB) levels using PixelCNN++ model. With the input of the original SMB maps data, we can predict the arrangement of different tiles for a brand new SMB level.

Blood cell counting is an essential part of the study of blood cells and hematopoiesis. Conven- tional methods require Hematology and hematopoiesis requires a hematology analyzer, other experimental equipment, and staining with chemicals, which is a time-consuming process. In this study, we used an object detection algorithm, ”you only look once” (YOLO) to automatically identify and count the unstained blood cells of African clawed frogs. We changed the dataset used for training. We compared the performance of the model by changing and enhancing the data set used for training. The results showed that the original dataset augmented by randomly changing the brightness had the highest accuracy. However, the accuracy of discriminating be- tween leukocytes and thrombocytes did not reach a practical level regardless of which dataset was used for training. This is due to the fact that the ratio of leukocyte and plug cell labeled objects to red blood cells is too low.

I introduced a novel multi-agent competition-based framework to train reinforcement learning agents to play Pokemon. Experiments showed that agents trained with out method outperformed those trained with self-play, even without reward shaping.

We proposed and experimented on multiple conditional Generative Adversarial Networks models for colorizing line art based on the given color hints by the user.

We defined an interface (gyms environment) to SUMO so that modeling of automated driving can be done using inter-vehicle communication and reinforcement learning, and verified that the interface works by conducting reinforcement learning experiments.

WaveNet has achieved high accuracy in the Text-To-Speech task. We hypothesized and experimented that it may perform well in chart-generation task.

I applied reinforcement learning to NetHack in order to learn strategies. I proposed the model for this purpose and confirmed its effectiveness.

We introduced LoL-V2T: a large video-to-text dataset and generated captions for esports footage with deep learning.

We propose a combination of 3 deep learning models to estimate the key which you want to press based on the hand poses which is taken with one RGB camera.

We proposed a method for learning efficiently when studying mathematics. Focusing on the dependency of the problem, we measured the efficiency of problem selection.

I worked on the experiment that creators can easily complete incomplete coloring which occurs while painting by deep learning.

I worked on the experiment that the model would be able to generate line image of characters from its illustration automatically using deep learning. Adding to typical loss function, I used 'perceptual loss' to raise accuracy of the model.

I developed Neural Networks for airfoil estimation, and the NN performed 1000x faster than conventional analysis. In addition, I developed web interface of the NN for airplane novices.