[Seminar] “On the way to the automated discovery of novel patterns and dynamics in physical and chemical systems by intrinsically motivated deep learning. ” by Dr. Reinke

Date

2019年5月30日 (木) 14:00 〜 15:00

Location

Seminar room B503, Level B, Lab1 Bldg.

Description

Dear all,

Neural Computation Unit (Doya Unit) would like to invite you to a seminar as follows.

Date: Thursday, May 30, 2019
Time: 14:00 – 15:00
Venue: Seminar room B503, Lab1 Bldg.

Speaker: Chris Reinke
                 Flowers Lab, Inria Bordeaux, France

Title:
On the way to the automated discovery of novel patterns and dynamics in physical and chemical systems by intrinsically motivated deep learning.
 
Abstract:
Many physical and chemical systems that we encounter are complex and dynamic. For example, the air flow around a rotor blade of a wind turbine might form different turbulence patterns over time. Those dynamics depend on parameters such as the form of the rotator blade, wind direction and others. To understand such systems we need to explore which dynamics and patterns can be produced. Unfortunately, many experiments might be needed to uncover those dynamics, which if manually done are often very time intensive. A solution comes from the application of AI and robotics, which perform and control the exploration of such systems automatically.

In our project, we are investigating the usage of intrinsically motivated goal exploration processes (IMGEPs) for the discovery of novel dynamics and patterns in physical and chemical systems. This new family of algorithms uses goal spaces about their target systems. The goal space defines goals that the agent might want to achieve, for example, to produce a certain turbulence pattern. The agent self-generates its own goals based on principles of intrinsic motivation that are inspired by human cognitive development (Deci and Ryan, 1985). The agent then explores and learns which system parameters achieve the generated goals. The framework has been shown to allow real world robots to efficiently acquire skills such as tool use in high-dimensional continuous state and action spaces. In the domain of chemistry and physics, they opened the possibility to automate the discovery of novel chemical or physical structures in oil droplet experiments (e.g. Grizou, Points et al., 2018). A focus of our project is to investigate how goal space representations for dynamical systems can be learned based on raw sensory observations, for example, video camera images.
 
During the seminar I will introduce the algorithms and their potential use for the discovery of novel dynamics and patterns in physical and chemical systems. Moreover, I will present first results for the discovery of complex self-organized visual patterns in Lenia, a continuous game-of-life cellular automaton.

 

We hope to see many of you at the seminar.

Sincerely,
Neural Computation Unit
Contact: ncus@oist.jp

All-OIST Category: 
セミナー

Intra-Group Category

Technical Seminar

Subscribe to the OIST Calendar: Right-click to download, then open in your calendar application.