Deep Learning: The Fundamentals

Fundamental concepts of a neural network

• Basic mathematics refresher.
• Neural network: Architectures, activation functions, and previous activation weights...
• Training a neural network: Cost functions, backpropagation, stochastic gradient descent...
• Modeling a neural network: Modeling input and output data according to the type of problem.
• Understanding functions with neural networks.
• Understanding distributions with neural networks.
• Data growth: How to balance the dataset?
• Generalization of neural network results.
• Initialization and regularization of the neural network: L1/L2 regularization, batch normalization.
• Optimization and convergence algorithms.

Demonstration: Adjustment functions and distributions using neural networks.

Common Machine Learning and Deep Learning tools

• Data management tools: Apache Spark, Apache Hadoop.
• Common Machine Learning tools: Numpy, Scipy, Sci-kit.
• High-level DL frameworks: PyTorch, Keras, Lasagne.
• Low-level DL frameworks: Theano, Torch, Caffe, Tensorflow. Demonstration
• Applications and limitations of the tools presented.

Convolutional Neural Networks (CNN)

• Fundamental principles and applications.
• Basic functioning of a CNN: convolutional layer, use of a kernel, padding and stride, etc.
• CNN architectures that have advanced the state of the art in image classification: LeNet, VGG Networks, Network in Network, etc.
• Use of an attention model.
• Application to a typical classification scenario (text or image).
• CNNs for generation: super-resolution, pixel-by-pixel segmentation.
• Main strategies for increasing feature maps for image generation.

Case study: Innovations brought about by each CNN architecture and their more global applications (1x1 convolution or residual connections).

Recurrent Neural Networks (RNN)

• Introduction to RNNs: fundamental principles and applications.
• Basic characteristics of RNNs: hidden activations, backpropagation in time, unfolded versions.
• Development for GRU (Gated Recurrent Units) and LSTM (Long Short Term Memory).
• Convergence problems and gradient leakage.
• Types of classic architecture: time series prediction, classification, etc. RNN encoder-decoder architecture. Use of feedforward models.
• NLP applications: word/character encoding, translation.
• NLP application: prediction of the next image generated from a video sequence.

Demonstration: Different states and developments brought about by Gated Recurrent Units and Long Short Term Memory architectures.

Generative models: VAE and GAN

• Presentation of Variational AutoEncoder (VAE) and Generative Adversarial Networks (GAN) generative models.
• Autoencoder: dimensionality reduction and limited generation.
• Variational AutoEncoder: generative model and approximation of data distribution.
• Definition and use of latent space. Reparameterization trick.
• Fundamentals of Generative Adversarial Networks. Convergence of a GAN and difficulties encountered.
• Fundamentals of Generative Adversarial Networks. Convergence of a GAN and difficulties encountered.
• Improved convergence: Wasserstein GAN, BeGAN. Earth Moving Distance.
• Applications for image or photograph generation, text generation, super resolution.

Demonstration: Applications of generative models and use of latent space.

Deep Reinforcement Learning

• Reinforcement Learning.
• Use of a neural network to understand the state function.
• Deep Q Learning: experience replay and application to video game control.
• Learning policy optimizations. On-policy and off-policy. Actor critic architecture. A3C.
• Applications: control of a simple video game or digital system.

Demonstration: Control of an agent in an environment defined by a state and possible actions.