Very grateful and honored to have been awarded an ERC Consolidator Grant of ~2 million Euro for my project GROW-AI: Growing Machines Capable of Rapid Learning in Unknown Environments.
#ERCCoG
Thanks to the
@ERC_Research
, my lab, and family for their support!
Introducing Neural Developmental Programs (NDPs)🧬🧠Instead of neural networks with fixed architectures, we allow neural networks to grow through a dynamic self-organizing process, inspired by how biological nervous systems develop👇
PDF:
Excited to share our work on Morphogenesis in Minecraft! We show that neural cellular automata can learn to grow not only complex 3D artifacts with over 3,000 blocks but also functional Minecraft machines that can regenerate when cut in half 🐛🔪=🐛🐛
PDF:
Finally done with my first blog post "The Future of Artificial Intelligence is Self-Organizing and Self-Assembling"!
Covering work from our group and others on the combination of ideas from deep learning and self-organizing systems.
Want to create your next game levels through natural language 🗣️🎮? Wait no more, we present:
"MarioGPT: Open-Ended Text2Level Generation through Large Language Models".
PDF:
MarioGPT also predicts the player's path!
Biological brains 🧠 are characterized by neurons of many different classes. Yet most artificial neural networks only use one activation function throughout.
We show that through neural diversity alone we can solve RL problems in networks with random weights 🤯
Tired of training neural networks day in and day out? Maybe train virtual creatures instead and join us for the next Virtual Creature Competition
@GECCO2019
Deadline: June 23rd w/
@Kriegmerica
and Nick Cheney. Check out past winners:
@enasmel
and myself are excited to announce our paper "Meta-Learning through Hebbian Plasticity in Random Networks"
Instead of optimizing the neural network's weights directly, we only search for synapse-specific Hebbian learning rules. Thread 👇
Many of the NeurIPS papers I'm currently reviewing:
"Our algorithm is great. Here are some graphs. If you want any details on how anything works, see our 40 page appendix." 🤔
Happy our "Deep Learning for Video Game Playing" review article is now published in IEEE Transactions on Games: PDF: Lots of new methods added since the last update!
@nojustesen
@togelius
@FilipoGiovanni
Our paper "Player-AI Interaction: What Neural Network Games Reveal About AI as Play" was accepted at
#CHI2021
!
Core finding: AI as play can expand current notions of human-AI interaction, which are currently predominantly productivity-based.
PDF:
This is the most detailed model of a human cell to date, obtained using x-ray, NMR and cryoelectron microscopy datasets.
“Cellular landscape cross-section through a eukaryotic cell.”
- by Evan Ingersoll and Gael McGill.
EvoCraft: A New Challenge for Open-Endedness!
Can we automatically search for increasingly complex artifacts/machines in an open-ended fashion in Minecraft?
PDF:
Python API:
First competition on open-endedness coming:
@GeccoConf
!👇
Happy to announce our new
@GECCO2019
paper on "Deep Neuroevolution of Recurrent and Discrete World Models" with
@kenneth0stanley
. A simple GA is able to learn the model end-to-end and finds effective representations for the visual and memory component.
Check out our Evolutionary Strategies (ES) Python library! Simple to use, support for different optimizers, PyTorch integration, ability to handle non-normal distributions, etc.
I'm super pleased to release EvoStrat
It's a library that makes Evolutionary Strategies (ES) simple to use.
I'm really happy how cleanly it separates the environment, the reinforcement learning agent, the population distribution and the optimization.
Congratulations to
@enasmel
for his (and my!) first accepted
#NeurIPS
paper! 🎉
We added an experiment that shows that Hebbian networks can also sometimes generalize to robot morphologies not seen during training.
PDF:
Code:
@enasmel
and myself are excited to announce our paper "Meta-Learning through Hebbian Plasticity in Random Networks"
Instead of optimizing the neural network's weights directly, we only search for synapse-specific Hebbian learning rules. Thread 👇
Backpropagation is nice and stuff but if we need to work so hard on making it not catastrophic forget or continually learn, maybe there is a better solution out there?
We introduce Evolve&Merge, a genomic bottleneck algorithm that iteratively reduces the no. of Hebbian rules in a network, learning which connections should share the same rules: 384 rules can self-organize random weights in a few steps to control a quadruped! w/
@JoachimWinther
Very excited to have been awarded a Google Faculty Research Award. Thank you very much for being my sponsor
@hardmaru
! More about the project soon but this summarises it pretty well 🤖👾⚕. Can't wait to get started!
@ITUkbh
Happy to announce our
#ICLR2022
paper: Variational Neural Cellular Automata! Lead by
@rasmusbergpalm
w/
@miguelgondu
,
@SudhakaranShyam
. We use an NCA as decoder for a VAE, allowing one VNCA to grow a variety of outputs purely through local communication.
Many don't know that Hypernetworks, NeRF, SIREN, etc. approaches all originate from a method first popularized in the evolutionary computation community: Compositional Pattern Producing Networks by
@kenneth0stanley
. If you haven't already, read:
I sometimes miss the days were we were not collectively hunting for the next state-of-the-art AI results but rather trying to gain deeper insights by studying algorithms on smaller (toy) problems... with the benefit of actually understanding what and how they learned.
Happy to present a new game we developed "iNNk: A Multi-Player Game to Deceive a Neural Network"
Paper:
Players need to communicate a secret code word to each other through drawings, without being deciphered by the neural network
Introducing Deep Innovation Protection - we scale neuroevolution to more complex tasks such as "VizDoom: Take Cover" by temporally reducing the selection pressure on specific components in an evolving world model. w/
@kenneth0stanley
@UberAILabs
@ITUkbh
It's pretty sad that some godfathers of deep learning are so dismissive of other ideas. Even if you don't believe in the ideas yourself, you shouldn't discourage others from pursuing them. Diversity is useful! It's ironic because they had to defend ANNs for so long themselves.
Inspired by how biological neural networks are grown through a self-organized developmental process, we introduce: HyperNCAs 🧬🧠 - Growing Developmental Networks with Neural Cellular Automata
PDF:
Led by
@EIiasNajarro
w/
@SudhakaranShyam
&
@claire__aoi
Presenting EPLS: Evolutionary Planning in Latent Space! Building on the world model approach by
@hardmaru
&
@SchmidhuberAI
, we use Rolling Horizon Evolution to find a sequence of continuous actions that maximize reward in a learned model of the world 👇
I got promoted to Full Professor in 2020 but waited with my inaugural lecture so my parents could attend. Just want to say thank you to everyb. that made this possible, especially current&past members of
@RoboEvoArtLab
! It's a privilege to work with these amazing people every day
Controversial opinion? Most current forms of RL "reinforce" a view in which learning and deployment are two separate phases, while there is nothing like that in biological systems. In biological systems learning never stops and we learn exclusively during deployment.
Today, I'm excited to announce Microsoft is teaming up with
@OpenAI
to exclusively license GPT-3, allowing us to develop and deliver advanced AI solutions for our customers.
#MSIgnite
We are happy to present "Regenerating Soft Robots through Neural Cellular Automata" w/
@khoribe3
,
@katt_walker
The approach allows soft robots to regrow parts of their morphology when being damaged only based on local cell communication.
PDF:
Congratulations to
@miguelgondu
! First accepted paper as PhD student!
We show that intelligent trial-and-error algorithms, originally developed for fast robot adaptation, can be used to find levels with the right difficulty in only a few trials.
w/
@rasmusbergpalm
,
@hardmaru
Deep learning systems are often brittle because they are trained and then frozen, while self-organizing systems in nature never stop adapting.
I'll present some of our ideas on Lifelong Learning in Self-Organizing Machines at the
#CVPR2022
Workshop on Continual Learning today!
Safe online reinforcement learning is still a major challenge. Here we introduce a novel approach: Instinct-regulated RL (IR²L).
An instinct network is pre-trained on a single task where it is safe to make mistakes and transferred to tasks in which learning safely is critical.
CLIP has been the backbone of many advances in ML. How vulnerable is CLIP to adversarial attacks?
Presenting CLIPMasterPrints, which are capable of maximizing the confidence score for a number of widely varying prompts, while being unrecognizable for humans🤯
Anybody outside DeepMind that has access to 🦩? Everybody from DeepMind posting impressive (cherry-picked?) successes of Flamingo and not many failure cases is surely a good way to fuel the AI hype.
Very much looking forward to giving a keynote at the
@GeccoConf
"Evolutionary Reinforcement Learning Workshop" about our recent work on scaling deep Neuroevolution to more complex tasks. Workshop takes place July 10th.
Interested in doing research on foundation models for video games? If you have a Master's degree with experience in ML and want to do an industrial PhD () together with
@ITUkbh
and
@modl_ai
, reach out! Please RT 👐
It's challenging to control self-organizing systems! We introduce an approach to learn to control them: Goal-Guided Neural CAs😃
Cell behavior is controlled dynamically at every step of cellular growth, enabling morphing and movement tasks. w/
@SudhakaranShyam
&
@EIiasNajarro
Excited to announce the 2nd edition of our Minecraft Open-Endedness Challenge
@GeccoConf
! Can you create an algorithm that produces a never-ending stream of increasingly complex artefacts?
More info:
Deadline: June 1st
Below highlights from last year.
Neural networks can be used for so much more than "just" machine learning. They offer novel game mechanics and even allow complete new types of games. Any game that uses neural networks in interesting ways, which we might have forgotten? w/
@jichenz
Excited to share our work on Morphogenesis in Minecraft! We show that neural cellular automata can learn to grow not only complex 3D artifacts with over 3,000 blocks but also functional Minecraft machines that can regenerate when cut in half 🐛🔪=🐛🐛
PDF:
Finally done with my first blog post "The Future of Artificial Intelligence is Self-Organizing and Self-Assembling"!
Covering work from our group and others on the combination of ideas from deep learning and self-organizing systems.
Interactive Evolution and Exploration Within Latent Level-Design Space of GANs
We introduce a tool for interactive GAN-based evolution of game levels. It also allows for direct exploration of the latent space, and allows users to play the levels.
PDF:
Neural architecture evolution, a new method for automatically finding optimal neural networks for video understanding, has yielded architectures that outperform existing hand-made models and show improvements to network runtime of 10-100x. Learn more ↓
So far confirmed sponsors for our Minecraft Open-Endedness Challenge are
@modl_ai
and
@OpenAI
!
You already know open-endedness is cool 😎 but if you need additional motivation to submit, we already have over $2000 USD in monetary awards 💸
More info:
Looking for a full-time or part-time (at least 20h/w) research assistant for 6 months to work on this open-endedness project. Knowledge of self-attention architectures and evolutionary algorithms a plus. Remote work possible. Start Jan/Feb. PM me with your CV if interested 🤓
EvoCraft: A New Challenge for Open-Endedness!
Can we automatically search for increasingly complex artifacts/machines in an open-ended fashion in Minecraft?
PDF:
Python API:
First competition on open-endedness coming:
@GeccoConf
!👇
EvoCraft: A New Challenge for Open-Endedness!
Can we automatically search for increasingly complex artifacts/machines in an open-ended fashion in Minecraft?
PDF:
Python API:
First competition on open-endedness coming:
@GeccoConf
!👇
Can humans outsmart machines to communicate a secret codeword visually? In our recent FDG paper we investigate how to deal with adversarial player strategies in our neural network game iNNk.
The video shows one such strategy, drawing a rebus puzzle.
PDF:
Starting to hire PhDs and postdocs on my GROW-AI
#ERCCoG
project:
Application deadline: February 1st, 2023.
Join us and get in touch if interested in how we can grow neural networks for more general agents! Aiming to build a diverse team. Please share.
Very grateful and honored to have been awarded an ERC Consolidator Grant of ~2 million Euro for my project GROW-AI: Growing Machines Capable of Rapid Learning in Unknown Environments.
#ERCCoG
Thanks to the
@ERC_Research
, my lab, and family for their support!
Congratulations to
@JoachimWinther
who successfully defended his Ph.D. thesis
@ITUkbh
🙌🎉! Joachim's exciting work has focused on bio-inspired approaches to adaptive agents. See some highlights below:
Come work with us! Multiple PhD and Postdoc positions available in my lab (deep RL, neuroevolution, quality diversity), thanks to generous funding from
@DFF_raad
More info here: Please share widely! We aim at building a diverse team. Deadline: April 15th.
Check out this great blog post, covering also some of our work on NCAs and morphogenesis!
Transformers are so 2022... collective intelligence for deep learning is where things are going 😉
New blog post: Collective Intelligence for Deep Learning
Recently,
@yujin_tang
and I published a paper about how ideas like swarm behavior, self-organization, emergence are gaining traction in deep learning.
I wrote a blog post summarizing the key ideas:
Our company is looking for a team member passionate about building world class game engine tools (Unity / Unreal Engine) around AI/ML. Come work with us! More information:
What happens when you slice a Minecraft caterpillar in half 😬 and try to regrow it with a neural cellular automata? Find out soon! Only virtual animals were harmed during these experiments.
If neuroscience needs evolution (as the authors here argue for convincingly), a more general artificial intelligence surely needs artificial evolution as well.
Very honored and humbled to having received the Distinguished Young Investigator award from the International Society for Artificial Life
@alifelab
couldn’t have done it without my great students, collaborators and previous advisors
Another crucial study showing how dangerous indoor situations can be. An alarming 94 (out of 216) people on the same floor in a call center in Korea got infected—an attack rate of 43.5%! But look at the schema. Most infections are on one side of the room.
Trying to write my first blog post, something like "A vision for a more self-organizing kind of Artificial Intelligence". Announcing it here so you can hold me accountable if I don't post it in a couple of days/(weeks).
Are your virtual creatures out of shape?
Start training them now so they’ll be ready in 8 weeks for the 2020 Virtual Creatures Competition
held virtually at
@ALifeConf
.
Submission deadline: June 26.
Looking forward to talking about our research on combining innate mechanisms and learning in evolving agents at this "Innateness in Neuroscience and AI" symposium and discussing the future of innateness w/ researchers such as
@MelMitchell1
,
@tdietterich
,
@AlisonGopnik
,
@mtkostecki
Great episode and great advice from
@geoffreyhinton
: "It's always a good idea to figure out what huge numbers of very smart people are working on and to work on something else."
"Population Based Training (PBT) is a method first developed at DeepMind that helps discover effective and efficient training regimes for neural nets."? I didn't know DeepMind invented evolutionary algorithms 🤔
We’re excited to announce the results of our research collaboration with
@Waymo
, applying Population Based Training (PBT) to help make the process of training neural nets in their self-driving cars more effective and efficient.
Our Glitch Finder bot searches for glitches in level geometry at 20 times the speed of human players.
Hear more about the possibilities in using AI and ML in game development at
@Devcom_conf
19 Aug 17:00-18:00 CEST (8am-9 PDT/11am-12 EDT)
#ddc2020
#gamedev
#AI
#GameAI
New paper on Evolving Scalable Neural Turing Machines. Trains on small bit vector sizes (9-bits) for a copy task and scales to 1000-bits without further training. PDF:
Most ANNs fail catastrophically if inputs are presented in an ordering unseen during optimization. In our recent
@GeccoConf
paper,
@JoachimWinther
and myself study what are the minimal network models to create input/output permutation-invariant agents.
Very excited to announce that our company
@modl_ai
has raised $8.4 million! Video game development and especially automated game testing will look very different with AI-driven bots 🤖
We are thrilled to announce that we’ve raised $8.4 million to redefine the game development process. We aim to make it more efficient and enjoyable.
We’re eager to get the product to you soon; this funding will help us get there faster.
We propose the popular board game Blood Bowl
as a new challenge for AI. The branching factor is several orders of magnitude larger than those of classic board games and scoring points in the game is rare and difficult, which makes it
hard to apply RL. PDF:
Our
#CoG2019
paper () presents the Bot Bowl competition, the motivation behind it, future plans, and preliminary results. So far, our RL agents can reach a medium playing level on a small game board with fixed teams and a limited set of starting formations.
Happy our review paper is published! We review the field of Game AI, which not only deals with creating agents that can play a certain game, but also with areas as diverse as creating game content automatically, game analytics, or player modelling. PDF:
Winner of the competition is: Nicholas Guttenberg (
@ngutten
)! Runner-ups are the Flowers Team (Etcheverry,
@mayalen_etc
, Chan
@BertChakovsky
, Moulin-Frier
@Clement_MF_
, Oudeyer
@pyoudeyer
) and the CIIRC/CVUT/INRIA team: Cisneros, Sivic, Mikolov. Congratulations!!
We got some great submissions for the first Minecraft Open-Endedness challenge! Here is a teaser of some of them. More this Sunday
@GeccoConf
. Winners will be announced at the end of the conference 👀
Very happy that our paper "CG-GAN: An Interactive Evolutionary GAN-based Approach for Facial Composite Generation" was accepted at AAAI-20. Huge congrats to my former
@ITUkbh
Masters students and lead authors Nicola Zaltron and Luisa Zurlo for this achievement!🤗
@GaryMarcus
@techreview
Maybe every paper abstract should have a mandatory field of what the limitations of the proposed approach are. That way some of the science miscommunications and hypes could maybe be avoided.
CPPN2GAN: Combining CPPNs and GANs for Large-scale Pattern Generation
We combine the ability of CPPNs to define latent
vector GAN inputs as a function of geometry, allowing large-scale pattern generation (e.g. Zelda levels). w/ Jacob Schrum and
@CIGbalance
I did not contribute, but this NEAT-Gym package from Simon Levy and Coletta Fuller (neither on Twitter) may be interesting to some. It lets you run Python-based NEAT/HyperNEAT/ES-HyperNEAT/novelty search in OpenAI Gym:
Open PhD positions funded by the Danish Data Science Academy. If you're interested in applying and doing a PhD at
@ITUkbh
, get in touch!
Call information:
Deadline: March 20th
3 PhD positions available: (1) GROW-AI project on developmental ANNs, (2) interdisciplinary grant on directing morphogenetic patterning in bio. and artificial systems, and (3) industrial PhD with on AI for game testing. If still at
#NeurIPS2023
, talk to me
Time flies! My first paper as a PhD student in 2009 with
@kenneth0stanley
investigated deceptiveness in meta-learning tasks. e.g. an agent has to learn to remember the position of a high reward in a T-maze environment. What do I mean with deceptiveness? 👇
There are 12 postdoc positions available in various fields (including explainable AI, AI in games, deep learning) as part of a collaboration between different Copenhagen universities. Please RT. Let's work together!
@ITUkbh
This self-organizing system is inherently robust to perturbations, enabling setting a large fraction of the cells’ states to zero and to recover almost perfectly by additional iterations of growth.
Updated "Increasing Generality in Machine Learning through Procedural Content Generation" paper now on . More figures, more citations, more tables!😃Thank you for comments on the previous version that definitely improved the paper. Did we miss anything?
It's time for reinforcement learning researchers to take the domains they train on seriously. Training on a set of fixed scenarios/levels leads to brittle policies that don't generalize. Procedural content generation can help, argue
@risi1979
and I here:
After a few rejections happy this work w/
@kenneth0stanley
is finally accepted and will appear at
#AAA2021
! Now just need to finally make the code pretty enough to be publicly releasable....
Introducing Deep Innovation Protection - we scale neuroevolution to more complex tasks such as "VizDoom: Take Cover" by temporally reducing the selection pressure on specific components in an evolving world model. w/
@kenneth0stanley
@UberAILabs
@ITUkbh
Earliest work where a neural network is used to create the architecture of a neural network (i.e. controller-based NAS)? We showed it can be done with ES-HyperNEAT in 2012, which determines both the architecture and weights of a neural network by an evolved hypernetwork.
New framework and paper: UnityVGDL, a port of the Video Game Description Language to Unity.
Code:
Paper:
We hope people make cool stuff with it! It integrates well with the
@unity3D
ML-agents.
This week I'll be giving a talk about UnityVGDL, a port of the GVGAI VGDL framework with
@unity3d
ML-Agents integration. The framework is available here:
Tired of the same mainstream machine learning conferences? If you're working at the intersection of evolutionary computation and ML, consider submitting to the Evolutionary Machine Learning track at
@GeccoConf
instead! Abstracts due: Jan 27th 🦎🧬🧠