Week 4 — Topify

This week, we discussed what deep learning algorithm should we use. We made research on various similar projects. But there weren’t many projects that used Deep Learning for this problem.

After our meeting with TA, we realized that a point that even though our dataset is balanced with the ratio of 4:1, we have few samples for the top tracks class. Therefore, we decided to look into N-Shot Learning examples.

Zero-Shot Learning was one of our options and it is used for the estimation of class samples that are not available, but One-Shot learning seemed like a more suitable option because we have samples for each class.

One-Shot Learning

One-shot learning is a challenge of object classification that is often seen in computer vision or NLP. But we will be using in Song Classification. Unlike another machine learning-based object classification algorithms, which involve hundreds or thousands of samples and very large datasets to practice, one-shot learning attempts to learn knowledge about object classes from only one or a few training samples.

We wanted to take a look at the simple examples of One-Shot learning that we see on the Internet and see how this learning works and what it can do.

This is a simple example of image classification with One-Shot Learning. For this example, Omniglot dataset is used. We observed that Siamese Neural Networks is usually used for One-Shot Learning. The network is learning a similarity function, which takes two images as input and expresses how similar they are.