Learning to compareIn the past, a concept called "moments" was found. With it, came the gift of shape matching without training, and the concept of object recognition was established, uniting the computer vision community in peace. Decades passed before deep learning was presented. Pain and confusion flowed in its wake, and people were struggling with training. This is a story of the revolution time. We know it, because the concept of moments has returned, and there is an algorithm that does not have to be retrained every time... (a parody to homeworld 2 story.) GitHub |
Learning to searchIn this work, I venture into the the learning to learn paradigm using a neural network. The result is a neural network that learns to search through the given input space to find a configuration that matches the provided figure. Learning to search can be considered an alternative to learning with less data, or even one-shot learning. It does not need any variation in the training example; because the model has to search for it with the possibility to extend to multi-object classification. GitHub |
Line scanning based OCRThis project utilizes the state-of-the-art tool to tackle one of the most classical problem in machine learning: optical character recognition. Unlike previous attempt in Thai language where users have to supply character breaks, the training paradigm of this project is end-to-end; namely, it takes a raw input image of words or sentences on a line, and automatically align each character to its corresponding output Unicode-label in the label list in maximum likelihood fashion. GitHub |
Shadow playAt Bit.Studio, we know that imagination is the best toy. This project aims to bring back our good old vivid memories with shadow play and re-enlivens it with captivating graphics. The shadow casted on the screen by player hands will be analyzed and morphed into a graphic animal that moves around, such imaginative play would never be outdated. Site |
A thinking machine"I think therefore I am." It has always been my life long endeavor to understand the thinking mechanism. I created this project as an empirical tool for the evaluation of my theorized computational model of thinking published earlier. GitHub PDF |
Emotion analyzerCan drawing strokes tell something about emotion? This project exploits the relationship between hand-painted strokes and painters' emotion to predict one from the other. Waiting for publicizing at Bit.Studio |
Decomposition based style replicatorStyle transfer is a very popular topic despite the best effort to improve the generation performance. Limited by the no-free-lunch theory, I tried to derive the mathematical relation between styles and locations of images resulting in a quick and ad-hoc style transfer machine powered by TensorFlow. This project has not yet been finalized; but when it is done, it will only requires to be trained once, no need to "charge" the network for new styles. BitBucket |
Solid light recognizerI brought neural network into a virtual reality to improve players experience. Using Unreal 4 engine, I could create a virtual world where players can use their hand gestures to draw magical runes in 3D space while the system, powered by TensorFlow, recognizes these and calls upon corresponding magics enchanting with spectacular visual effect. Waiting for publicizing at Bit.Studio |
Conditional generative modelThis is my very first project with TensorFlow. I went back to verify the idea of conditional data generation where users can specific what category of data they want to see. The network can generate data not only correspondingly to users' behests but also conformingly to the given styles. Site |
Orientation sorterAs a tutorial for deep learning training for my friends, I tried to show that the performance of a neural network can be improved with unsupervised layer pre-training. I recorded images of my own hands with Kinect2 , and adapted a python script with TensorFlow that describes a neural network that can sort the orientation of my hand in the images. Site |