Jekyll2026-10-27T18:44:55+01:00https://benlonnqvist.github.io/news/feed.xmlBen LonnqvistBen Lonnqvist is Head of NeuroAI at Read/Write Labs. His research compares human and machine vision at scale and builds brain models precise enough to act on.Ben Lonnqvistbenhlonnqvist [at] gmail [dot] comJournal of Neural Engineering: deep-learning features improve simulated prosthetic vision2026-08-01T00:00:00+02:002026-08-01T00:00:00+02:00https://benlonnqvist.github.io/news/2026/jne-prostheticNew paper with Elsa Scialom, Alban Bornet and Michael Herzog in the Journal of Neural Engineering. We use deep networks to predict which image features are diagnostic for recognising an object. Rendering those features in simulated prosthetic vision improves how well people recognise objects.

]]>Ben Lonnqvistbenhlonnqvist [at] gmail [dot] comICLR 2026: Model-Guided Microstimulation Steers Primate Visual Behavior2026-04-23T00:00:00+02:002026-04-23T00:00:00+02:00https://benlonnqvist.github.io/news/2026/iclr-microstimOur paper with Johannes Mehrer, Anna Mitola, Paolo Papale and Martin Schrimpf was presented at ICLR 2026.

We pair a model that predicts behaviour from both image input and cortical stimulation with a procedure that maps the model’s best stimulation sites onto monkey visual cortex. In two macaques, model-guided microstimulation steered their object choices. It is a proof of principle for visual prostheses that target higher-level visual cortex.

]]>Ben Lonnqvistbenhlonnqvist [at] gmail [dot] comPhD thesis: Differences in Visual Perception in Humans and Deep Neural Networks2025-09-05T00:00:00+02:002025-09-05T00:00:00+02:00https://benlonnqvist.github.io/news/2025/phd-defenseI publicly defended my PhD thesis, Differences in Visual Perception in Humans and Deep Neural Networks, at EPFL. The thesis was advised by Michael Herzog and Martin Schrimpf.

]]>Ben Lonnqvistbenhlonnqvist [at] gmail [dot] comIn the press: AI can’t see as well as humans, and how to fix it2025-07-25T00:00:00+02:002025-07-25T00:00:00+02:00https://benlonnqvist.github.io/news/2025/contour-pressMy work has recently been covered by the press.

Graphical abstract

EPFL Press

My paper on a large-scale comparison of humans and DNNs has been picked up by media:

Nik Papageorgiou from the EPFL press wrote a fantastic article on my work which can be read for free here.

In addition, EPFL covered the work on its front page! Certainly a proud and humbling note to finish my Ph.D. with, having privately defended my thesis successfully just a couple weeks prior. Thank you EPFL!

EPFL Front Page

Swiss Press

Local Swiss media also picked up the article. 20 Minutes, a Swiss French-speaking magazine read by over 400,000 people daily published an article on the work, which you can read for free here.

]]>Ben Lonnqvistbenhlonnqvist [at] gmail [dot] comICML 2025: Contour Integration Underlies Human-Like Vision2025-07-15T00:00:00+02:002025-07-15T00:00:00+02:00https://benlonnqvist.github.io/news/2025/contour-icmlOur new paper on a large-scale comparison of humans and DNNs has been presented at ICML 2025! We show that many models fail when object contours are fragmented, and that the ability to do contour integration is critical for performance and robustness.

Graphical abstract

Resources

]]>Ben Lonnqvistbenhlonnqvist [at] gmail [dot] comIn the press: Latent Noise Segmentation in Tekniikka&Talous2025-02-13T00:00:00+01:002025-02-13T00:00:00+01:00https://benlonnqvist.github.io/news/2025/lns-pressMy paper on Latent Noise Segmentation, published at the top conference ICML, has recently been picked up by the Finnish press. I gave an interview to Finnish journalist Raili Leino at EPFL.

The article was published in the Finnish magazine Tekniikka&Talous, the largest Finnish magazine on technical subjects. In fact, the magazine reaches over 500,000 people annually.

You can read the fantastic article on my work here.

Link to paper: https://proceedings.mlr.press/v235/lonnqvist24a.html

]]>Ben Lonnqvistbenhlonnqvist [at] gmail [dot] comCCN 2024 poster: current DNNs can’t integrate information across object discontinuities2024-08-09T00:00:00+02:002024-08-09T00:00:00+02:00https://benlonnqvist.github.io/news/2024/ccnI attended Cognitive Computational Neuroscience in Boston this year and presented my new work on how Current DNNs are Unable to Integrate Visual Information Across Object Discontinuities.

We also have a more up-to-date poster, which you can download at this link.

]]>Ben Lonnqvistbenhlonnqvist [at] gmail [dot] comICML 2024: Latent Noise Segmentation2024-07-23T00:00:00+02:002024-07-23T00:00:00+02:00https://benlonnqvist.github.io/news/2024/lns-icmlOur paper on: Latent Noise Segmentation has been accepted at ICML 2024! We show that neural noise can be used for unsupervised segmentation, and doing so reproduces many hallmark results of human perceptual grouping.

Link to paper: https://proceedings.mlr.press/v235/lonnqvist24a.html

]]>Ben Lonnqvistbenhlonnqvist [at] gmail [dot] comNew preprint: Latent Noise Segmentation2023-09-29T00:00:00+02:002023-09-29T00:00:00+02:00https://benlonnqvist.github.io/news/2023/latent-noise-segmentationCheck out our new pre-print: Latent Noise Segmentation! We show that neural noise can be used for unsupervised segmentation, and doing so reproduces many hallmark results of human perceptual grouping.

Link to pre-print: https://arxiv.org/abs/2309.16515

]]>Ben Lonnqvistbenhlonnqvist [at] gmail [dot] com