\[
\begin{aligned}
\text{if } & \Phi((x, y)) = x \text{ and } f((x, y)) = y \\
\text{ then } & f((3, 1)) \neq f((3, 2)) \\
\text{ but } & \Phi((3, 1)) = \Phi((3, 2)) \\
\end{aligned}
\]
Separate Signal and Noise
Data Augmentation: Same Class
Data Augmentation: Noise
Feature Extraction
Convolutional Neural Network
Goal of the Thesis
Goal: Improve upon the state-of-the-art baseline
New convolutional neural network architecture
New data augmentation technique
Meta-data fusion
Question: Can these techniques be used to improve classification accuracy?
Network architecture: deep residual neural networks
Introduce shortcuts
Allow training of very deep networks
Shortcut
\[ y = \mathcal{H}(x) = \mathcal{F}(x) + x \]
Data augmentation: Multiple-width frequency-delta
Image source: Yoonchang Han and Kyogu Lee. Acoustic scene classification using convolutional neural network and multiple-width frequency-delta data augmentation. 14(8):1–11, 2016