Understanding Deep Neural Networks Training Course

Part 1 – Deep Learning and DNN Concepts

Introduction to AI, Machine Learning & Deep Learning

• History, fundamental concepts, and common applications of artificial intelligence, beyond the myths often associated with the field
• Collective Intelligence: aggregating knowledge shared by numerous virtual agents
• Genetic algorithms: evolving a population of virtual agents through selection
• Conventional Machine Learning: definition
• Types of tasks: supervised learning, unsupervised learning, reinforcement learning
• Types of actions: classification, regression, clustering, density estimation, dimensionality reduction
• Examples of Machine Learning algorithms: Linear regression, Naive Bayes, Random Tree
• Machine Learning vs Deep Learning: problems where Machine Learning remains the state of the art (Random Forests & XGBoosts)

Basic Concepts of a Neural Network (Application: multi-layer perceptron)

• Recap of mathematical foundations
• Definition of a neural network: classical architecture, activation functions, and
• Weighting of previous activations, network depth
• Definition of neural network learning: cost functions, back-propagation, stochastic gradient descent, maximum likelihood
• Modelling a neural network: modelling input and output data according to the problem type (regression, classification, etc.). The curse of dimensionality
• Distinguishing between multi-feature data and signals. Choosing a cost function based on the data
• Function approximation using a neural network: presentation and examples
• Distribution approximation using a neural network: presentation and examples
• Data Augmentation: how to balance a dataset
• Generalisation of neural network results
• Initialisation and regularisation of a neural network: L1/L2 regularisation, Batch Normalisation
• Optimisation and convergence algorithms

Standard ML / DL Tools

A concise overview covering advantages, disadvantages, ecosystem positioning, and usage is planned.

• Data management tools: Apache Spark, Apache Hadoop
• Machine Learning: NumPy, SciPy, scikit-learn
• High-level DL frameworks: PyTorch, Keras, Lasagne
• Low-level DL frameworks: Theano, Torch, Caffe, TensorFlow

Convolutional Neural Networks (CNN)

• Overview of CNNs: fundamental principles and applications
• Basic operation of a CNN: convolutional layers, kernel usage
• Padding & stride, feature map generation, pooling layers. Extensions to 1D, 2D, and 3D
• Overview of different CNN architectures that achieved state-of-the-art performance in classification
• Images: LeNet, VGG Networks, Network in Network, Inception, ResNet. Presentation of innovations introduced by each architecture and their broader applications (e.g., 1x1 convolutions or residual connections)
• Use of attention models
• Application to a common classification case (text or image)
• CNNs for generation: super-resolution, pixel-to-pixel segmentation. Presentation of
• Main strategies for increasing feature maps in image generation

Recurrent Neural Networks (RNN)

• Overview of RNNs: fundamental principles and applications
• Basic operation of an RNN: hidden activation, back-propagation through time, unfolded version
• Evolution towards Gated Recurrent Units (GRUs) and LSTM (Long Short-Term Memory)
• Overview of different states and the advancements brought by these architectures
• Convergence and vanishing gradient problems
• Classical architectures: time series prediction, classification, etc.
• RNN Encoder-Decoder architecture. Use of attention models
• NLP applications: word/character encoding, translation
• Video applications: prediction of the next generated image in a video sequence

Generative Models: Variational AutoEncoder (VAE) and Generative Adversarial Networks (GAN)

• Overview of generative models and their connection to CNNs
• Auto-encoder: dimensionality reduction and limited generation
• Variational Auto-encoder: generative model and distribution approximation of a given dataset. Definition and use of latent space. Reparameterisation trick. Applications and observed limitations
• Generative Adversarial Networks: fundamentals
• Dual network architecture (generator and discriminator) with alternating learning and available cost functions
• GAN convergence and encountered difficulties
• Improved convergence: Wasserstein GAN, Began, Earth Mover's Distance
• Applications for image or photograph generation, text generation, and super-resolution

Deep Reinforcement Learning

• Overview of reinforcement learning: controlling an agent within a defined environment
• Via states and possible actions
• Using a neural network to approximate the state function
• Deep Q-Learning: experience replay and application to video game control
• Policy optimisation. On-policy && off-policy. Actor-critic architecture. A3C
• Applications: control of a single video game or a digital system

Part 2 – Theano for Deep Learning

Theano Basics

• Introduction
• Installation and configuration

Theano Functions

• Inputs, outputs, updates, givens

Training and Optimisation of a Neural Network using Theano

• Neural network modelling
• Logistic regression
• Hidden layers
• Training a network
• Computing and classification
• Optimisation
• Log loss

Testing the Model

Part 3 – DNN using TensorFlow

TensorFlow Basics

• Creating, initialising, saving, and restoring TensorFlow variables
• Feeding, reading, and preloading TensorFlow data
• Using TensorFlow infrastructure to train models at scale
• Visualising and evaluating models with TensorBoard

TensorFlow Mechanics

• Prepare the data
• Download
• Inputs and placeholders
• Build the graphs
  • Inference
  • Loss
  • Training
• Train the model
  • The graph
  • The session
  • Training loop
• Evaluate the model
  • Build the evaluation graph
  • Evaluation output

The Perceptron

• Activation functions
• The perceptron learning algorithm
• Binary classification with the perceptron
• Document classification with the perceptron
• Limitations of the perceptron

From the Perceptron to Support Vector Machines

• Kernels and the kernel trick
• Maximum margin classification and support vectors

Artificial Neural Networks

• Non-linear decision boundaries
• Feedforward and feedback artificial neural networks
• Multilayer perceptrons
• Minimising the cost function
• Forward propagation
• Back propagation
• Improving the way neural networks learn

Convolutional Neural Networks

• Goals
• Model architecture
• Principles
• Code organisation
• Launching and training the model
• Evaluating a model

Brief introductions will be provided for the following modules (subject to time availability):

TensorFlow – Advanced Usage

• Threading and queues
• Distributed TensorFlow
• Writing documentation and sharing your model
• Customising data readers
• Manipulating TensorFlow model files

TensorFlow Serving

• Introduction
• Basic serving tutorial
• Advanced serving tutorial
• Serving Inception model tutorial