1

What is Artificial Intelligence (up to Neural Networks)?

• The myth of artificial intelligence versus today’s reality.
• Intellectual tasks versus algorithms.
• Types of actions: classification, regression, clustering, density estimation, dimensionality reduction.
• Collective intelligence: aggregating shared knowledge from numerous virtual agents.
• Genetic algorithms: evolving a population of virtual agents through selection.
• Machine learning: an overview and key algorithms (XGBoost, Random Forest).

2

Neural Networks and Deep Learning

• What is a neural network?
• What does it mean for a neural network to learn? Deep vs. shallow networks, overfitting, underfitting, convergence.
• Understanding a function through a neural network: presentation and examples.
• Generation of internal representations within a neural network.
• Generalization of a neural network’s results.
• The deep learning revolution: the generality of tools and challenges.

3

Applications of Deep Learning

• Data classification: various scenarios including raw data, images, audio, text, etc.
• The challenges of data classification and the decisions involved in choosing a classification model.
• Classification tools: using Multilayer Perceptrons or Convolutional Neural Networks in machine learning.
• Forecasting and sequential/temporal data: challenges and limitations of information prediction.
• Structural rules within data that can enable predictive logic; common forecasting tools.
• Data transformation/generation: reinterpretation operations such as denoising, image segmentation…
• Transformation within the same format: translating text from one language to another.
• Generation of “original” data: neural style transfer, generating images from textual descriptions.
• Reinforcement learning: controlling an environment.
• Experience replay and training neural networks on video games.

4

What Problems Can Be Solved with Machine/Deep Learning?

• Data conditions: volume, dimensionality, class balance, description.
• Raw data versus engineered features: making the right choice.
• Machine learning versus deep learning: using traditional ML algorithms or neural networks.
• Defining the problem: Unsupervised Learning versus Supervised Learning.
• Assessing the solution: understanding the gap between an assertion and the output of an algorithm.

5

Preparing a Dataset

• Defining a dataset and distinguishing it from a traditional database (DB).
• Storing and monitoring data: controlling biases, cleaning/converting while allowing for iterative improvements.
• Understanding the data: using statistical tools to analyze data distribution and representation.
• Formatting data: deciding on input and output formats, linking them to the problem definition.
• Preparing the data: defining the Training Set, Validation Set, and Test Set.
• Establishing a framework to ensure that the algorithms used are truly relevant (or not).

6

Searching for the Optimal Solution

• Methodologies for progressing towards the best solution for an ML/DL problem.
• Choosing a research direction, locating similar publications or existing projects.
• Iterating from the simplest algorithms to the most complex architectures.
• Maintaining a comprehensive benchmarking framework.
• Achieving an optimal solution.

7

The Tools

• What tools are available today?
• Which tools are best suited for research versus industrial applications?
• From Keras/Lasagne to Caffe, Torch, Theano, TensorFlow, Apache Spark, or Hadoop.
• Industrializing a neural network by strictly managing its process and ensuring continuous monitoring.
• Implementing successive retraining sessions to keep a network current and optimal.
• Training users to understand the neural network.