Online Sequential Optimization

Course from O. Wintenberger for Master M2A @ Sorbonne Université.
Website: http://wintenberger.fr/ens.html

Implemented Algorithms

• Gradient Descent, Constrained Gradient Descent
• Stochastic Gradient Descent, Constrained SGD
• Regularized Follow the Leader (RFTL) variants: Stochastic Mirror Descent (SMD), Stochatic Exponential Gradient (SEG), Adaptative Gradient (AdaGrad)
• Bandits: SREG, SBEG
• ADAM variant: Adam, Projected Adam, AdaMax, Adam Temporal, Adamax Temporal

Getting Started

The MNIST data set can be downloaded from http://yann.lecun.com/exdb/mnist/
The notebook main.ipynb contains a sandbox for all the implemented algorithm for the user to experiment with the hyperparameters and compare respective performances.
All the algorithms are benchmarked in a binary classification problem, tuning a linear SVM for model.

Default config results

Contributors

Lise Le Boudec, Paul Liautaud, Nicolas Olivain