Accuracy is a lie
Evolution spent billions of years creating intelligence. I'm compressing that into days. I evolve AI rather than train it: self-organizing, self-improving networks that keep surprising me. They find solutions I did not design and would not have thought of, and since none of it is fitted to human data, there is no reason for the result to look familiar. Working out whether that is genuinely new or only looks new to us is the part I am still on.
Evolution Over Engineering
Where everyone sees failure, I see evolution at work • Building AI that assembles components instead of memorizing patterns, because adaptive behavior beats accuracy.
What I Actually Do
Evolutionary AI: Growing Intelligence Instead of Training It
While everyone's training networks the usual way, I'm evolving behaviors. Slower? Yes. More compute? Absolutely. But mine adapt when yours break.
Research Through Systematic Exploration
Used more CPU hours than sensible. Set a record along the way, but the real signal was the behaviors carrying over to new problems on their own. That's when you know you're onto something fundamental.
Compositional Intelligence: AI That Thinks in Parts
My networks solve problems by assembling solutions from components. Not elegant. Not efficient. But interpretable and adaptable.
Research Philosophy
Building AI that adapts instead of memorizing
Current Focus
What I am working on now, and the problem sitting under it.
Why Evolution
Why I grow networks instead of training them.
My Approach
How I work, mostly learned by wasting time first.
Seeking
What I want from collaborators and from problems.
Selection Pressure
From engineering precision to evolving intelligence • I once thought I had answers; now I realize I only have questions. Each pivot stripping away what I thought I knew, teaching me that growth happens at the edge of understanding. The pattern matches evolution: survival goes to whatever adapts, right or wrong. Now I let each experience prove that everything is possible when you embrace mutation over optimization.
University Teaching Specialist
Doctoral Assistant
Visiting Researcher
Projects
Loading projects...
The Evolution Lab
Started trying to simulate the brain. Ended up evolving intelligence • Every field I touched taught me the same lesson: minds have to grow. Engineering them top-down never worked for me.
My Research Journey
Most researchers find their field. I had a question that wouldn't let me settle: child me wanted thinking robots. Pursued that dream through physics, mechanics, neuroscience, AI, until they all revealed the same truth: minds evolve into existence. The irony? Decades of education had turned me into the robot, trained to engineer things. I can't escape that mindset, but now I'm breeding artificial life into existence, watching behavior arise from chaos rather than from architecture, code, or engineering. Whether that is where intelligence actually begins is the question, not the conclusion.
GEENNS: Compositional Intelligence Through Evolution
Lifetime Research Project (Post-PhD Phase)
Intelligence is not one big network. It is a collaboration of specialists that evolution grows, freezes, and learns to recombine. GEENNS pushes that to its hard edge: get the collaboration itself to emerge, so behavior comes from evolution rather than from copying us or a dataset. The PhD proved the parts can be frozen and reused. Making the whole emerge is the work now.
Compositional
A mind as a team of specialists: split a problem into parts, hand each to a network that grew for it, then combine their answers
The bet is easy to say and hard to earn: intelligence is a collaboration of specialists, and the collaboration itself has to be evolvable. A problem gets split, handed to networks that each grew for a piece of it, and their answers composed into one. The proof of concept already does this, evolved specialists solving compound problems that a single network fails.
Evolutionary
Grown, not engineered: evolution discovers the specialists and how they cooperate, so behavior emerges instead of being coded
I design nothing by hand. Evolution grows each specialist's wiring, and evolution grows the coordinators that decide how the specialists work together. My job is to set the conditions and the pressure, then watch what the search finds. The structure and the cooperation are discovered, not authored.
Growing
The live frontier: roles that specialists claim on their own, coordination that carries to unseen problems, wholes that outdo their parts
Each specialist is a small evolved circuit repeated across a grid, inspired by the cortex without copying it: one template, many behaviors, depending on how it gets wired in. The roadmap lets these substrates grow and adapt like living tissue rather than sit as fixed graphs, so a network can develop and keep changing after it is born.
Lifelong
Freeze and reuse: settled specialists stay fixed while new tasks only add coordination, a path to learning across a lifetime without forgetting
For a new task the specialists stay frozen and only the coordination evolves, so hard-won skills are reused instead of retrained. That is the road to learning across a lifetime without overwriting the past. The honest status: today's composition is still task-specific, but the signal for something more general is already there in the population, and pulling it out into genuine emergent reasoning is the open problem I am on.
Will be public upon publication
Emerging Behaviors: Intelligence Without a Human Template
Active Research Direction
Networks that are grown rather than trained never see human data, so what they do has no obligation to resemble anything we would recognize. I study that on its own terms. The open problem is telling behavior that is genuinely new from behavior that only looks new to us, because every measure we have is calibrated on human data.
Scaling Adaptive Substrate Neuroevolution
PhD Thesis
Grow a big network from a tiny recipe, the way DNA grows a body. It never worked in practice. My thesis diagnoses why, then redesigns it to run at real scale.
GraphQA: Engineer Intelligence to Think Slow
Master's Thesis
Built zero-shot conversational AI using sub-knowledge graphs. It proved to me that engineering creates brittle intelligence.
Overclouds: When Privacy Met Democracy
Bachelor's Thesis
Built anonymous, decentralized data sharing right through the browser. It taught me that distributed systems stand or fall on trust.
When Senses Collide: Visual-Vestibular Integration
Pre-Undergrad Research
Watched brains fuse conflicting senses into truth. My first proof that intelligence comes from integrating senses rather than selecting one.
The Paper Trail
Leaving breadcrumbs of a longer journey • Documenting discoveries that captured what I knew at the time. Looking back, they were all converging.
Peer-Reviewed Publications
Journal articles and conference proceedings
2026
Multi-Behavioral Evolved Substrates Through Neuromodulation and Activation Selection
Romain Claret, Michael O'Neill, Paul Cotofrei, Kilian Stoffel
Artificial Life Conference (ALIFE)
Neuromodulation alone stalls at a 75% parity ceiling; adding per-task activation selection lets one evolved genotype solve all five tasks at once, on every seed.
2026•3 citations
Tensor-Accelerated Eager Multi-Resolution Grids for Evolving Large-Scale Substrates
Romain Claret, Michael O'Neill, Paul Cotofrei, Kilian Stoffel
Genetic and Evolutionary Computation Conference (GECCO)
EMR-HyperNEAT recasts substrate discovery as one batched tensor pass, reaching 12 to 34x per-generation GPU speedups on XOR and validating evolution up to depth 13.
2026
Bio-Inspired Palette Evolution in Indirectly Encoded Substrates: Timescale Compatibility Shapes Activation Function Discovery
Romain Claret, Michael O'Neill, Paul Cotofrei, Kilian Stoffel
Parallel Problem Solving from Nature (PPSN)
Eleven bio-inspired strategies for evolving activation palettes converge up to twice as fast as a tuned baseline, with success shaped by timescale compatibility.
2026•2 citations
Per-Node Activation Function Evolution in Indirectly Encoded Substrates: Solvability, Limits, and Emergent Diversity
Romain Claret, Michael O'Neill, Paul Cotofrei, Kilian Stoffel
Artificial Life Conference (ALIFE)
Evolving per-node activation functions reveals a Parity-4 divide (oscillatory 100%, monotonic 0%) specific to evolutionary search in sparse substrates.
2026
Breaking the Central Bias: Spatially Partitioned Experts for Coordinate-Based Neuroevolution
Romain Claret, Arthur Gygax, Michael O'Neill, Paul Cotofrei, Michael Palma Mendes, Pascal Felber
International Conference on Pattern Recognition (ICPR)
Splitting MNIST across 13 evolved spatial experts breaks ES-HyperNEAT's central-pixel bias: pixel coverage jumps from 4% to 79% and mean accuracy more than doubles to 43%.
2026
Early-Stopping Thresholds for ES-HyperNEAT: A Data-Driven Approach from Fitness Dynamics
Romain Claret, Arthur Gygax, Michael O'Neill, Paul Cotofrei, Pascal Felber
IEEE World Congress on Computational Intelligence (WCCI)
A data-driven early-stopping rule keeps over 90% of successful ES-HyperNEAT runs on MNIST, cuts compute by 41.6%, and is 64% more efficient than Hyperband.
2024•5 citations
Investigating Hyperparameter Optimization and Transferability for ES-HyperNEAT: A TPE Approach
Romain Claret, Michael O'Neill, Paul Cotofrei, Kilian Stoffel
Genetic and Evolutionary Computation Conference (GECCO)
Achieved 29% MNIST accuracy with ES-HyperNEAT through systematic TPE optimization, beating the previous 23.90% benchmark and transferring successfully to Fashion-MNIST.
Other Work
Technical reports, white papers, and additional contributions
2026
On Scaling Coordinate-Based Neuroevolution: The Quadtree Bottleneck in ES-HyperNEAT
A JAX GPU implementation (JAX-ESHN) pinpoints why ES-HyperNEAT resists scaling: each CPPN discovers its own substrate positions, blocking population-level vectorization.
2023
Neuroevolution to Simultaneously Execute Heterogeneous Tasks
Scaling neuroevolution for heterogeneous tasks using a universal knowledge representation.
2016
Blockchain, a techie overview
Demystifying blockchain when everyone thought it would change everything. Technical reality vs. religious fervor.
2010
Perceptual roll tilt thresholds demonstrate visual-vestibular fusion
Investigating how the brain integrates visual and vestibular information for motion perception by comparing precision thresholds across sensory modalities.
Degrees
Every degree was a detour that turned out to be the destination • Spent a decade in universities learning how to engineer. Had to master building before I could discover growing.
Thesis: Scaling Adaptive Substrate Neuroevolution
Supervisors: Prof. Dr. Kilian Stoffel and Adjunct Prof. Dr. Paul Cotofrei
MSc in Engineering
HES-SO - University of Applied Sciences and Arts Western Switzerland, Lausanne, SwitzerlandThesis: Multi-hop Multi-turns Question-Answering Chatbot using Sub-Knowledge Graphs
Supervisor: Prof. Dr. Jean Hennebert
BSc in Software Engineering
HE-Arc - Haute Ecole Arc, Neuchâtel, SwitzerlandThesis: Anonymous and Decentralized Browser-based Data Sharing Service
Supervisor: ing. info. dipl. EPF Marc Schaefer
Professional Certifications & Training
Side Quests That Mattered. Formal certifications and training programs that taught me a different way to approach problems.
Thoughts & Tales
The noise that doesn't fit in a paper • Research insights, creative fiction, technical rants, philosophical spirals. Some polished, most not.
Contact
Always collecting pieces of the puzzle • Especially the ones that don't fit. Research collaborations, wild theories, or proof I'm wrong all welcome.
