Imagine you are running on the piano and create beautiful songs. What a fantastic thing it will be? Now your dream come true, you can run and play with music. Let's start with a simple song...
Web

Although pianists may change one or several notes from the score when they are performing, whether these changes are noticeable to the audience still needs to be explored. Previous studies showed that non-diatonic changes can be easier to detect by adults than diatonic changes. In this study, we extended this topic to musicians and non-musicians by comparing their note change detection abilities under several conditions. We performed melody discrimination tests which contain the melodies with a strong tonal center and a weak tonal center, where each type of the melody contains a chord note change or a non-chord note change. We hypothesize that for both musicians and non-musicians, the note changes in a melody with a strong tonal center would be easier to detect than the melody with a weak tonal center, and the non-chord related change is easier to detect than chord related change. Results show that the musicians have a better performance in all the note change detection tasks than the non-musicians. However, there are no significant performance differences between the tasks using melodies with a strong tonal center and the tasks using melodies with a weak tonal center. Moreover, musicians can easily detect non-chord note changes than chord note changes, while non-musicians show no significant detection differences between the two types of changes overall.
Report

Spatial Audio Painting is a painting board which is correlated to the sound variations in space. When you paint on the canvas, the spatial position of the audio source will be changed as you paint in different locations.
Web

The physical modeling approach provides semantic sound representations for the parameters having physical interpretations. In this paper, piano modeling approach is surveyed and discussed. A typical piano modeling structure is presented. First the excitation part, then the string and radiation modeling. A multi-rate soundboard model is specifically discussed towards the high computational complexity of physical modeling.
Report

Natural language inference is a fundamental task in natural language understanding. Because of the understanding required to assess the relationship between two sentences, it can provide rich, generalized semantic representations. In this study, we implement a sentence-encoding model using recurrent neural networks. Our hypothesis was that semantic role labels, along with GloVe word embeddings, would give the sentences rich representations, making the model not only more successful on the original SNLI challenge, but also more robust to adversarial examples. However, our findings show that adding the SRL information does not improve the performance of our baseline model on either the SNLI task or the adversarial data sets.
Report

Recent studies in the field of structural engineering raise eyes on the importance of post-earthquake recovery in urban areas. With the modern remote sensing technology, building and structural component images become accessible via aerial drones or related devices. In this project, the state-of-the-art deep learning technology for a civil engineering application is implemented, namely recognition of structural damage from images. In the PEER PHI Kaggle Competition, we rank at 15% among other competitors.
Poster