Networks, Space and AI

with a grain of toxicity

Marcus Märtens

Follow this presentation on your device: https://coolrunning.github.io/esaentry_pres

Network
Architectures and
Services

Barrett Lyon / The Opte Project
Visualization of the routing paths of the Internet.

B. C. Coutinho, Sungryong Hong, Kim Albrecht, Arjun Dey, Albert-László Barabási, et al.
The Network Behind the Cosmic Web.

The hairball
Your everyday network.

Network Science

The study of the structure of local interactions between agents in large-scale systems.

Data-driven and inspired by the emergence of large-scale networks.
Multi-disciplinary in nature.
Theory based on statistical physics, mathematics and algorithmics.

Powerlaw degree distribution

$\Pr[k] \sim k^\gamma$

Scale-free Network

The Automated Network Scientist

Network scientists compute properties of networks.
Idea: Take properties as features and learn their relations!
Symbolic Regression: Find formulas for difficult to compute properties!

Idea: Evolve formula by Cartesian Genetic Programming (CGP)

Example: Network Diameter by Laplacian Eigenvalues

Genetic Programming can help to uncover hidden relations between network features!

a sample network

M. Märtens, P. Van Mieghem, F. Kuipers
Symbolic Regression on Network Properties, EUROGP2017, pp 131-146

Information flow in brain networks

MEG measurements of 78 brain regions
Compute Phase Transfer Entropy (PTE) to describe information exchange
Binarize matrix with threshold $\tau$ to get network
Network Motifs: Frequently occuring subnetworks

Clustering of the brain

Search for all motif frequencies in the network
Find statistically overrepresented motif
Find partition of nodes such that least number of motifs are cut
Map the clusters back to the brain

Spatially coherent brain regions suggest higher order organization of information processing!

M. Märtens, J. Meier, A. Hillebrand, P. Tewarie, P. Van Mieghem
Brain network clustering with information flow motifs, Applied Network Science, 2:25, Springer Open

Toxicity Detection in Multiplayer Games

Toxicity

Anti-social behavior towards own team-mates
Offensive language and insults
Data-driven study of chat communication of 10305 matches of Dota

Data full of insults

Toxicity Study

Deployed algorithms

Toxicity detection by $n$-grams
Supervised Learning: SVM, predict winning team
Unsupervised Learning: Topic Model, LDA
Toxicity as transfer entropy

Key findings

Win-rate largely unaffected by toxicity
Kill events are likely triggers of toxicity
Once chances of winning diminish, toxicity increases
Susceptibility to toxicity is strongly heterogeneous

M. Märtens, S. Shen, A. Iosup and F. Kuipers
Toxicity Detection in Multiplayer Onlien Games, IEEE NetGames2015, pp. 1-6

Let us talk about the future

Where is AI now?

human competitive performance at Go

human competitive at diagnosis of skin cancer

prototype of autonomous cars on the roads

questionable performance at towel folding?

Where does AI need to go?

Digital evolution

First computers had their programs hard-coded
Later ones could be reprogrammed by patching cables and pulling levers
Only after stored-program architectures were invented, computers became the versatile general purpose devices that they are today.

AI needs to do the same transition: away from hand-crafted systems towards general purpose architectures!

Key technologies

One-shot learning: Learning by few training examples
Transfer learning: Condense knowledge and apply accross domains
(Neuro)-evolution: Growing and adapting to new environments

First results in General Video Game Playing (GVGP)

Testbed for reinforcement learning, planning, transfer, strategic reasoning and causation

Stephen Kelly, Malcolm I. Heywood
Emergent Tangled Graph Representations for Atari Game Playing Agents, EUROGP17, (pp. 64-79), Springer
Ken Kansky, Tom Silver, et al.
Schema Networks: Zero-shot Transfer with a Generative Causal Model of Intuitive Physics, https://arxiv.org/abs/1706.04317

Proposals

Proposal I: Take part (and win) a GVGP competition!

Candidates: www.gvgai.net, OpenAI gym
Evaluate approaches and ACT tools (e.g. pagmo2, pyaudi, dCGP)
Development of expertise
Fun and glory

Proposal II: Organize an own (one-shot) competition

Usage of Kelvins platform: https://kelvins.esa.int
Exclusive data required (to avoid participants to multi-shot)
Space related theme
Potential challenges: control (landing), classification of earth observation images (disaster)