Deep learning is a booming machine-learning technique which we often read in a lot of articles nowadays. Deep learning sounds like an intimidating concept for a lot of people, but everyone believes that deep learning is a cutting- edge tool to solve a lot of problems. In this talk, we will see how Python and various open-source tools are very easy to use and very powerful for solving deep learning problems. For the study case, we will have a 30-minute journey in revisiting image recognition problem with anime characters.
I will briefly explain how traditional machine learning works with TensorFlow and introduces some alternatives tools out there. After that, we will see how deep learning enhances our knowledge in the traditional machine learning and how we can get more benefits from it. All of these examples will be presented in the context of image recognition problem, and while on it, we will also see the use cases of other tools, such as OpenCV, OpenFace, etc.
In the main part of the talk, we will see that images of anime characters are limited by nature, since it totally depends on number of existing fan arts. In addition, 3D model (human face) cannot be applied directly in solving this problem since 2D model (anime character face) totally depends on the illustrator. Therefore, we will explore how we can utilize transfer learning for model training with small amount of dataset. Finally, we will build our own simple app with Sanic to serve our ready-built model for other users.
This talk will introduce the concept of deep learning, transfer learning, and how we can apply it to our own existing problems. In addition, we will present some alternative PaaS as a comparison to build a model from scratch.