Hannes Mehrer
banner
hannesmehrer.bsky.social
Hannes Mehrer
@hannesmehrer.bsky.social
Computational neuroscientist, NeuroAI lab @EPFL
Model perceptual changes via simulated microstimulation

We visualized perceptual changes from simulated stimulations in model face-selective regions. This results in face-related changes: additional faces appear (#1), face becomes larger (#1161), or specific face-features get enhanced (#533).
October 7, 2025 at 3:22 PM
Experiment 2

With a slightly different site-selection criterion, stimulation shifted behavior above baseline in monkey 1 (Cohen’s d=0.67), though our model was not able to accurately predict monkey behavior anymore.
October 7, 2025 at 3:22 PM
Experiment 1

Model-predicted behavioral shifts correlated with stimulation-evoked behavioral shifts in both monkeys. While predicted model responses were strong, monkey behavior was not shifted above baseline.
October 7, 2025 at 3:22 PM
Visual stimuli via GANs

We generate image sequences that smoothly modulate neural activity along a stimulation site’s tuning dimension. This links visual input to the direction of activation changes resulting from microstimulation (Papale et al. 2024: www.biorxiv.org/content/10.1...)
October 7, 2025 at 3:22 PM
How it works

1. Map the in-silico cortical sheet of a topographic model to the monkey cortex.
2. Optimize stimulation parameters by prototyping experiments in the model.
3. Only test those parameters in-vivo that are predicted to yield the largest behavioral effects.
October 7, 2025 at 3:22 PM
Stimulate high-level vs early visual cortex

Visual prosthetics in early visual areas can evoke simple percepts (letters), but they are limited by 1. electrode count and 2. low-level features. We target high-level cortex to elicit percepts of more complex objects.
October 7, 2025 at 3:22 PM
Model perceptual changes via simulated microstimulation

We visualized perceptual changes from simulated stimulations in model face-selective regions. This results in face-related changes: additional faces appear (#1), face becomes larger (#1161), or specific face-features get enhanced (#533).
October 7, 2025 at 2:37 PM
Experiment 2

With a slightly different site-selection criterion, stimulation shifted behavior above baseline in monkey 1 (Cohen’s d=0.67), though our model was not able to accurately predict monkey behavior anymore.
October 7, 2025 at 2:37 PM
Experiment 1

Model-predicted behavioral shifts correlated with stimulation-evoked behavioral shifts in both monkeys. While predicted model responses were strong, monkey behavior was not shifted above baseline.
October 7, 2025 at 2:37 PM
Visual stimuli via GANs

We generate image sequences that smoothly modulate neural activity along a stimulation site’s tuning dimension. This links visual input to the direction of activation changes resulting from microstimulation (Papale et al. 2024: www.biorxiv.org/content/10.1...)
October 7, 2025 at 2:37 PM
How it works

1. Map the in-silico cortical sheet of a topographic model to the monkey cortex.
2. Optimize stimulation parameters by prototyping experiments in the model.
3. Only test those parameters in-vivo that are predicted to yield the largest behavioral effects.
October 7, 2025 at 2:37 PM
Stimulate high-level vs early visual cortex

Visual prosthetics in early visual areas can evoke simple percepts (letters), but they are limited by 1. electrode count and 2. low-level features. We target high-level cortex to elicit pecepts of more complex objects.
October 7, 2025 at 2:37 PM
Also from Amsterdam (Christina Sartzetaki, Gemma Roig, Cees G.M. Snoek, Iris Groen): Using RSA, 100 models are evaluated on the bold moments dataset (doi.org/10.1038/s414..., data available on openneuro: openneuro.org/datasets/ds0...).
April 28, 2025 at 3:17 AM
Similar to the @bashivan.bsky.social 2019 paper on neural population control (doi.org/10.1126/scie...), a team from Amsterdam (Diego Cerdas, Christina Sartzetaki, Magnus Petersen, Pacal Mettes, Iris Groen)
April 28, 2025 at 3:17 AM
Smoothly varying maps can also emerge in vision (note, here not! language) models through credit-based self-organizing maps.
April 28, 2025 at 3:17 AM
Another form of a topographic (language) model: TopoNets.
Here, the spatial loss is based on the difference between a layer's unit activations and a blurred (/low-pass filtered) version of it.
April 28, 2025 at 3:17 AM
Before ICLR 2025 comes to an end today, a few #NeuroAI impressions from Singapore.
First, very happy to present our work on TopoLM as an oral, here with
@neilrathi.bsky.social
initial thread: bsky.app/profile/hann...
paper: doi.org/10.48550/arX...
code: github.com/epflneuroailab
April 28, 2025 at 3:17 AM
Together with
@neil_rathi
I will present our #ICLR2025 Oral paper on TopoLM, a topographic language model!

Oral: Friday, 25 Apr 4:18 p.m. (session 4C)
Poster: Friday, 25 Apr 10 a.m. --> Hall 3 + Hall 2B Paper: arxiv.org/abs/2410.11516
Code and weights: github.com/epflneuroailab
April 23, 2025 at 5:30 AM