Probabilistic Inference & Control (and lots of open source) - - PowerPoint PPT Presentation

▶

Dec 29, 2022 265 likes •456 views

Probabilistic Inference & Control (and lots of open source) Patrick van der Smagt Director of AI Research Volkswagen Group Data Lab Munich CNNs proved to excel in supervised image classification great results from deep learning with

SLIDE 1

Probabilistic Inference & Control (and lots of open source)

Patrick van der Smagt Director of AI Research Volkswagen Group Data Lab Munich

SLIDE 2

CNNs proved to excel in supervised image classification

SLIDE 3

great results from deep learning with Q-learning

SLIDE 4

Kacelnik / Auersperg / von Bayern

U. Oxford / U Vienna

SLIDE 5

can you pick out the tufas?

from Josh Tenenbaum, but I first saw it from Nando de Freitas

“tufa” “tufa” “tufa”

SLIDE 6

"understand" what you see

SLIDE 7

Deep Variational Bayes Filtering: filtering in latent space of a variational autoencoder

p(z(t)) p(z(t+1))

system state x(t) system state x(t+1)

filter

system state x(t+2)

p(z(t+2))

system state x(t+1) system state x(t+2) ~ ~ ~ ~ system state x(t) ~

back- propagation

filter

unsupervised

Karl & Soelch & Bayer & van der Smagt, ICLR 2017

SLIDE 8

DVBF: coding in latent space inverted pendulum dynamics learned from images

Karl & Soelch & Bayer & van der Smagt, ICLR 2017

unsupervised

SLIDE 9

control through DVBF: exploration

SLIDE 10

control through DVBF: after unsupervised learning with Empowerment

unsupervised

maximises information gain

C(s) := max

p(a|s)

Z p(a|s) Z p(s0|s, a) ln p(s0|s, a) p(s0|s) ds0da

KL-divergence between

= max

p(a|s)

Z p(a|s) KL ⇥ p(s0|s, a) k p(s0|s) ⇤

Erwin Schrödinger, 1944: Negentropy Klyubin et al, 2005: Empowerment Wissner-Gross et al, 2013: Causal Entropic Forces

Karl et al, AISTATS 2018

SLIDE 11

we're reaching out

www.argmax.ai
Maximilian Karl, Maximilian Soelch, Justin Bayer, Patrick van der Smagt (2017)

Deep Variational Bayes Filters: Unsupervised Learning of State Space Models from Raw Data  International Conference on Learning Representations (ICLR)

Baris Kayalibay, Grady Jensen, Patrick van der Smagt (2017)

CNN-based Segmentation of Medical Imaging Data  arXiv

Nutan Chen, Maximilian Karl, Patrick van der Smagt (2016)

Dynamic Movement Primitives in Latent Space of Time-Dependent Variational Autoencoders 

Proc. 16th IEEE-RAS International Conference on Humanoid Robots
collaboration in European projects and with TUM, DLR, TU Berlin, U Berkeley, NVIDIA, DeepMind, U Edinburgh, U Freiburg, LMU,

Umeå University, ...

open-sourcing our sofuware, to promote its deployment and further development

SLIDE 12

"unrelated" open-source project: sigma—learning how computers learn

Florian Cäsar and Michael Plainer HTL Wels, Austria https://sigma.rocks/ https://github.com/ThinkingTransistor/Sigma

SLIDE 13

"unrelated" open-source project: sigma—learning how computers learn

GPU CPU #neurons per layer CPU GPU

SLIDE 14

Sigma awards

Best Project - National ITS Project Award 1st Prize
Best Science Project - National JugendInnovativ 1st Prize
EIROforum Special Donated Prize - International European Union Contest for Young Scientists
3rd Prize - International European Union Contest for Young Scientists

SLIDE 15

the 2017 Deep Learning and Robotics Challenge

rganised by Volkswagen Group AI Research w/ NVIDIA support

31 blog posts 21 students 10 Mindstorms 10 Jetson TX-2 5 teams 5 1080's 4.5 weeks 1 DGX-1 1 task: unsupervised brick sorting many CNN's, RNN's, GMM's, VAE's, Kalman filters, SLAM, neural networks, PID, kNN, ...

SLIDE 16

2017 Deep Learning and Robotics Challenge

SLIDE 17

2017 Deep Learning and Robotics Challenge the winning team: Great Dolphins Kaboom

Jonathon Luiten Lucia Seitz David Adrian Karolina Stosio Akshat Tandon

SLIDE 18

join us

http://argmax.ai
World Summit AI, Amsterdam, Oct. 11–12
Munich City AI: http://munich.city.ai on Oct. 26