In this post we implemented both the front-end and the back-end of Sesame Movie website utilizing React.js and Spring Boot respectively. The basic functions implemented in this website include browsing movies by trending or genres, browsing game detailed information, and user register or login to record their preferences on each movie. More functions like user creating movie lists, rate or write reviews for a movie.

AIMA is short for the book “Artificial Intelligence: A Modern Approach” which is written by Stuart Russell and Peter Norvig. Currently they are working on the 4th edition which will be published soon. In the new edition, there are many major changes to the book structure and machine learning algorithms. I worked in the open-source organization AIMA to implement the new and modified algorithms in Python. The algorithms in the book can fall into two categories: classic machine learning algorithms and deep learning algorithms.The initial proposal listed the plan of implementation of main algorithms in detail.

In this post, we proposed a two-stream wavelet transformation network based on the popular two-stream convolutional networks. Before going into the concept of the two-stream wavelet network, we will going through an application of how it can be used on CNNs in the area of action recognitions in videos. The key idea is to feed two streams of data into two separate CNNs: spatial stream that models scene and object contexts, while temporal-stream likely provides some motion-based attentions on foreground actions.

In this post we introduced the implemention of a encode-decoder model for picture captioning. The training dataset we utilized is the coco dataset which contains both the image and its caption extracted by human. We also used resnet-152 to extract the features of images and embedding layers to lower the dimension of text space. A simple LSTM cell class is also implemented to be used in the text interpretation.

This is a kaggle chanllenge launched by Santander to help looking for ways to help customers understand their financial health and identify which products and services might help them achieve their monetary goals. On the other hand, the chanllenge is about to help identify which customers will make a specific transaction in the future, irrespective of the amount of money transacted.

Style transfer is an interesting task to do, as it contains the basic ideas in image processing which is popular in model mobile apps. It is also a project that worth stressing in computer vision classes. In this post, we implemented a style transfer CNN on the top of the pre-trained VGG-16 or VGG-19 models. To achieve the goal of style transfer, two loss functions are utilized to optimize the neural network: the style loss and the content loss. By adjusting the ratio of these two loss functions, the user is able to control the degree of style transference.

Any tranformaton involving frequency domain operations is suitable for image processing. An image can be treated as a two-dimensional signal which can be applied many image processing techniques. High dimensional Fourier Transformation, Convolution and even Wavelet Transformation are suitable to extract certain features from a 2D signal with possibilities of various filters.

Wavelet transform is frequently used in picture denoising as it can cover a broader length in the frequency domain. Wavelet transforms are based on small wavelets with limited duration. Gabor filter is one of the most commonly used filter in the field of signal analysis.