Deep Learning to See

Abstract

Motion is the protagonist of visionThere's not a morning I begin without a thousand questions running through my mind ... The reason why a bird was given wings If not to fly, and praise the sky ...

Neural Time-Reversed Generalized Riccati Equation

“…In many instances of control problems, where either the complexity of the model is high and/or the temporal horizon could be very long, as it could happen for instance in Reinforcement Learning (Sutton and Barto 2018;Bertsekas 2019) or Lifelong Learning (Betti et al 2022;Mai et al 2022), these methods are unfeasible and we usually need to resort to different control strategies. A typical approach is that of using Model Predictive Control (Garcia, Prett, and Morari 1989) (also known as receding horizon control), with a real-time iteration (RTI) scheme for solving the online optimization problem (Diehl, Bock, and Schlöder 2005).…”

“…The Thirty-Eighth AAAI Conference on Artificial Intelligence This approach has been inspired by the possibility of using optimal control techniques in the continual online learning scenario recently proposed in (Betti et al 2022) to formulate a class of lifelong problems using the formalism of control theory. For this reason, throughout the paper we assume that the dynamical system that defines the evolution of the state is also expressed by a neural model, in the form of a continuous time recurrent neural network (Zhang, Wang, and Liu 2014).…”

“…For this reason, throughout the paper we assume that the dynamical system that defines the evolution of the state is also expressed by a neural model, in the form of a continuous time recurrent neural network (Zhang, Wang, and Liu 2014). The authors in (Betti et al 2022) proposed a method to enforce stability by pushing the costate dynamic to converge to zero, and hence directly interfering with the dynamics prescribed by Hamilton equations. Conversely, we directly leverage on Hamilton equations to devise a stabilizing policy for the system.…”