Applied mathematics doctoral course with Neural Networks and Deep Learning Approach
This course is built by professor Renato Souza and includes several topics:
- Supervised Machine Learning: Concepts and Process Workflow
- Neuron and Layers
- Universal approximation theorem
- Forward and BackPropagation, Gradient Checking
- Activation Functions
- Bias, Variance, Overfitting, Underfitting, Regularization
- Optimization, Batch and Mini-Batch
- Frameworks: PyTorch / Tensorflow
- Types of Neural networks
My partner Lucas Resck and I concluded the course with a work on Deep generative models for data creation (Deep fakes), exploring its first use in porn celebrities and how it became famous after Barack Obama's video was published.
After that, its usage only increased, with applications from entertainment videos and audios to politics.
- Encoder-Decoder Networks: Four neural networks are used alone or combined to create this kind of media: Two networks with narrower layers towards the center so that there is a encode in the latent space.
- Convolucional Neural Networks: Learn filters that move through the entrance forming a map of abstract features.
- Generative Adversarial Networks: A dispute between a neural network that generates fakes and a neural network that discriminates fakes from real ones.
-
Recurrent Neural Networks: Handles sequential variables. After processing
$x^{(i-1)}$ the network remembers the internal state and can use$x^{(i)}$ .
Nowadays, CycleGan is a great tool used to deal with audio input.