Software

ilqc
Toolbox for analyzing non-linear control algorithms from an optimization perspective.
– Implementation of classical nonlinear algorithms: gradient, Gauss-Newton, Newton, Differentiable Dynamic Programming approach with quadratic or linear-quadratic approximations.
– Oracles implemented à la Pytorch, various line-searches tested for all algorithms.
– Diverse environments available with animations: pendulum, pendulum on a cart, simple model of a car, bicycle model of a cart.
– Model Predictive Controller implemented for a contouring objective of a track racing task on a complex track.
conference paper follow-up (theory) follow-up (implementation) notebook tutorial

tpri
Implementation of Target Propagation oracles for Recurrent Neural Networks (RNNs). Our implementation uses the analytical formulation of the inverse of the layers of a RNN rather than an inverse approximated by an auto-encoder. Our experimental results demonstrate the potential of Target Propagation on several tasks for very long RNNs.
paper

xsdc
Implementation of semi-supervised learning methods with deep network feature representations (XSDC). Our work extends the discriminative clustering approach of F. Bach and Z. Harchaoui DIFFRAC to nonlinear feature representations. Our results show the benefits of incorporating nonlinear feature representation learning in semi-supervised learning in a flexible framework that can tackle any level of supervision in the data.
paper

casimir
Toolbox of selected optimization algorithms for unstructured tasks such as binary classification, and structured prediction tasks such as visual object localization and named entity recognition.
paper

yesweckn
Code to implement Convolutional Kernel Networks (CKNs) on image classification tasks. CKNs are the translation of Convolutional Neural Networks (ConvNets) into the framework of kernel methods. The present code provides a concrete implementation of this framework on benchmark architectures and datasets.
paper