Yacine Chitour

Cyprien Tamekue, Dario Prandi, Yacine Chitour

Understanding sensory-induced cortical patterns in the primary visual cortex V1 is an important challenge both for physiological motivations and for improving our understanding of human perception and visual organisation. In this work, we focus on pattern formation in the visual cortex when the cortical activity is driven by a geometric visual hallucination-like stimulus. In particular, we present a theoretical framework for sensory-induced hallucinations which allows one to reproduce novel psychophysical results such as the MacKay effect (Nature, 1957) and the Billock and Tsou experiences (PNAS, 2007).

Mehdi Benallegue, Rafael Cisneros, Abdelaziz Benallegue et al.

In this paper, we present an observation scheme, with proven Lyapunov stability, for estimating a humanoid's floating base orientation. The idea is to use velocity aided attitude estimation, which requires to know the velocity of the system. This velocity can be obtained by taking into account the kinematic data provided by contact information with the environment and using the IMU and joint encoders. We demonstrate how this operation can be used in the case of a fixed or a moving contact, allowing it to be employed for locomotion. We show how to use this velocity estimation within a selected two-stage state tilt estimator: (i) the first which has a global and quick convergence (ii) and the second which has smooth and robust dynamics. We provide new specific proofs of almost global Lyapunov asymptotic stability and local exponential convergence for this observer. Finally, we assess its performance by employing a comparative simulation and by using it within a closed-loop stabilization scheme for HRP-5P and HRP-2KAI robots performing whole-body kinematic tasks and locomotion.

Yacine Chitour, Zhenyu Liao, Romain Couillet

In this paper, we propose a geometric framework to analyze the convergence properties of gradient descent trajectories in the context of linear neural networks. We translate a well-known empirical observation of linear neural nets into a conjecture that we call the \emph{overfitting conjecture} which states that, for almost all training data and initial conditions, the trajectory of the corresponding gradient descent system converges to a global minimum. This would imply that the solution achieved by vanilla gradient descent algorithms is equivalent to that of the least-squares estimation, for linear neural networks of an arbitrary number of hidden layers. Built upon a key invariance property induced by the network structure, we first establish convergence of gradient descent trajectories to critical points of the square loss function in the case of linear networks of arbitrary depth. Our second result is the proof of the \emph{overfitting conjecture} in the case of single-hidden-layer linear networks with an argument based on the notion of normal hyperbolicity and under a generic property on the training data (i.e., holding for almost all training data).

Mehdi Benallegue, Abdelaziz Benallegue, Yacine Chitour

The paper presents a new observer for tilt estimation of a 3-D non-rigid pendulum. The system can be seen as a multibody robot attached to the environment with a ball joint. There is no sensor for the joint position of the sensor. The estimation of tilt, i.e. roll and pitch angles, is mandatory for balance control for a humanoid robot and all tasks requiring verticality. Our method obtains tilt estimations using encoders on other joints and inertial measurements given by an IMU equipped with tri-axial accelerometer and gyrometer mounted in any body of the robot. The estimator takes profit from the kinematic coupling resulting from the pivot constraint and uses the entire signal of accelerometer including linear accelerations. Almost Global Asymptotic convergence of the estimation errors is proven together with local exponential stability. The performance of the proposed observer is illustrated by simulations.