humans vs. AI
NeuroAI benchmarking
Large-scale tests of AI models against human behaviour and the brain, to find exactly where today's models stop seeing like we do. Over a thousand models compared.
7 papers →Ben Lonnqvist
Head of NeuroAI · Read/Write Labs
I build AI models of the brain precise enough to predict and create therapies, with the long-term goal of curing brain dysfunctions.
About
I'm Head of NeuroAI at Read/Write Labs (formerly Dandelion Science), where we use computational models of the brain to develop targeted therapies for brain disorders. I architect the company's generative neuromodulation platform.
I did my PhD in Computational Neuroscience at EPFL (2020–2025), advised by Michael Herzog and Martin Schrimpf. My thesis, Differences in Visual Perception in Humans and Deep Neural Networks, combined psychophysics, neural data and model-building. I'm also a core contributor to the open-source Brain-Score platform. Before that I studied Economics and Finance at the University of Aberdeen.
Research
Visual Grouping
Grouping & segmentation
How the visual system binds fragments into objects. Latent Noise Segmentation shows that neural noise alone can drive unsupervised segmentation.
3 papers →model → stimulation
Neurotechnology & prosthetics
Using brain models to decide where and how to stimulate, from model-guided microstimulation in primates to better prosthetic vision.
2 papers →Selected publications
All publications →2026
Model-Guided Microstimulation Steers Primate Visual Behavior
International Conference on Learning Representations
A model that predicts how image input and cortical stimulation jointly drive behavior is used to choose stimulation sites in macaque visual cortex. Model-guided microstimulation steered the monkeys' object choices, a proof of principle for prostheses that target higher-level visual cortex.
@inproceedings{mehrer2026microstimulation,
title = {Model-Guided Microstimulation Steers Primate Visual Behavior},
author = {Mehrer, Johannes and Lonnqvist, Ben and Mitola, Anna and Papale, Paolo and Schrimpf, Martin},
booktitle = {International Conference on Learning Representations (ICLR)},
year = {2026}
}ICLR
2025
Contour Integration Underlies Human-Like Vision
International Conference on Machine Learning (PMLR 267)
Humans and more than 1,000 models were tested on objects whose contours are progressively fragmented. Humans stay accurate with very few visible contours, while most models fall to near chance. Only models trained on around 5 billion examples approach human performance, and the human-like bias for directional fragments grows with training data.
@inproceedings{lonnqvist2025contour,
title = {Contour Integration Underlies Human-Like Vision},
author = {Lonnqvist, Ben and Scialom, Elsa and Gokce, Abdulkadir and Merchant, Zehra and Herzog, Michael H. and Schrimpf, Martin},
booktitle = {Proceedings of the 42nd International Conference on Machine Learning},
series = {Proceedings of Machine Learning Research},
volume = {267},
pages = {40290--40311},
year = {2025}
}ICML
2024
Latent Noise Segmentation: How Neural Noise Leads to the Emergence of Segmentation and Grouping
International Conference on Machine Learning (PMLR 235)
Neural noise in a variational autoencoder's latent space is enough to segment images without supervision. The resulting model reproduces many hallmark effects of human perceptual grouping.
@inproceedings{lonnqvist2024latent,
title = {Latent Noise Segmentation: How Neural Noise Leads to the Emergence of Segmentation and Grouping},
author = {Lonnqvist, Ben and Wu, Zhengqing and Herzog, Michael H.},
booktitle = {Proceedings of the 41st International Conference on Machine Learning},
series = {Proceedings of Machine Learning Research},
volume = {235},
year = {2024}
}ICML
News
All news →- Journal of Neural Engineering: deep-learning features improve simulated prosthetic vision
- ICLR 2026: Model-Guided Microstimulation Steers Primate Visual Behavior
- PhD thesis: Differences in Visual Perception in Humans and Deep Neural Networks
- In the press: AI can't see as well as humans, and how to fix it
- ICML 2025: Contour Integration Underlies Human-Like Vision
Contact
For research collaborations, talks or press, contact me on LinkedIn.