Machine Learning Project-Papers reviews

Abstract

Introduction In my final project, I chose five papers related to machine learning of how to train the neural network with low audio resource data. The reason I chose this subject is because we already learned how to analyze numerical data, and training images in CNN model, but didn’t learn the audio learning. Also, in my opinion, the audio classification has much future prospect to research. It’s the most common way for people to communicate with others, if the machine learning can learn well on the audio classification, the technology will significantly affect the future trending on the smart devices. The other reason is the data collecting would be more difficult in the future. Since the privacy becomes more important for the society, how to train the model on low resource data will be very important. According to above reasons, I chose “Training neural audio classifiers with few data” [1] as the main paper. It used prototypical networks and transfer learning to train the neural network. Other four papers are cited by this paper [1], also related to audio training in low-data resources. In these five papers, some used the same data set (UrbanSound8K), and others used different set (ASC-TUT, Neural Information Processing Scaled for Bioacoustics (NIPS4B) bird song competition of 2013). However, collecting audio data is expensive, and time-consuming for the human to manually labeled. How to train the neural network prediction accurately is the main point.