Tim Kietzmann
banner
timkietzmann.bsky.social
Tim Kietzmann
@timkietzmann.bsky.social
ML meets Neuroscience #NeuroAI, Full Professor at the Institute of Cognitive Science (Uni Osnabrück), prev. @ Donders Inst., Cambridge University
Friday keeps on giving. Interested in representational drift in macaques? Then come check out Dan's (@anthesdaniel.bsky.social) work providing first evidence for a sequence of three different, yet comparatively stable clusters in V4.

Time: August 15, 2-5pm
Location: Poster C142, de Brug & E‑Hall
August 8, 2025 at 2:21 PM
Another Friday feat: Philip Sulewski's (@psulewski.bsky.social) and @thonor.bsky.social's
modelling work. Predictive remapping and allocentric coding as consequences of energy efficiency in RNN models of active vision

Time: Friday, August 15, 2:00 – 5:00 pm,
Location: Poster C112, de Brug & E‑Hall
August 8, 2025 at 2:21 PM
Also on Friday, Victoria Bosch (@initself.bsky.social) presents her superb work on fusing brain scans with LLMs.

CorText-AMA: brain-language fusion as a new tool for probing visually evoked brain responses

Time: 2 – 5 pm
Location: Poster C119, de Brug & E‑Hall
2025.ccneuro.org/poster/?id=n...
August 8, 2025 at 2:21 PM
On Friday, Carmen @carmenamme.bsky.social has a talk & poster on exciting AVS analyses. Encoding of Fixation-Specific Visual Information: No Evidence of Information Carry-Over between Fixations

Talk: 12:00 – 1:00 pm, Room C1.04
Poster: C153, 2:00 – 5:00 pm, de Brug &E‑Hall
www.kietzmannlab.org/avs/
August 8, 2025 at 2:21 PM
On Tuesday, Sushrut's (@sushrutthorat.bsky.social) Glimpse Prediction Networks will make their debut: a self-supervised deep learning approach for scene-representations that align extremely well with human ventral stream.

Time: August 12, 1:30 – 4:30 pm
Location: A55, de Brug & E‑Hall
August 8, 2025 at 2:21 PM
First, @zejinlu.bsky.social will talk about how adopting a human developmental visual diet yields robust, shape-based AI vision. Biological inspiration for the win!

Talk Time/Location: Monday, 3-6 pm, Room A2.11
Poster Time/Location: Friday, 2-5 pm, C116 at de Brug & E‑Hall
August 8, 2025 at 2:21 PM
Result 4: How about adversarial robustness? DVD-trained models also showed greater resilience to all black- and white-box attacks tested, performing 3–5 times better than baselines under high-strength perturbations. 8/
July 8, 2025 at 1:04 PM
Result 3: DVD-trained models exhibit more human-like robustness to Gaussian blur compared to baselines, plus an overall improved robustness to all image perturbations tested. 7/
July 8, 2025 at 1:04 PM
Result 2: DVD-training enabled abstract shape recognition in cases where AI frontier models, despite being explicitly prompted, fail spectacularly.

t-SNE nicely visualises the fundamentally different approach of DVD-trained models. 6/
July 8, 2025 at 1:04 PM
Layerwise relevance propagation revealed that DVD-training resulted in a different recognition strategy than baseline controls: DVD-training puts emphasis on large parts of the objects, rather than highly localised or highly distributed features. 5/
July 8, 2025 at 1:04 PM
Result 1: DVD training massively improves shape-reliance in ANNs.

We report a new state of the art, reaching human-level shape-bias (even though the model uses orders of magnitude less data and parameters). This was true for all datasets and architectures tested 4/
July 8, 2025 at 1:04 PM
We then test the resulting DNNs across a range of conditions, each selected because they are challenging to AI: (i) shape-texture bias, (ii) recognising abstract shapes embedded in complex backgrounds, (iii) robustness to image perturbations, and (iv) adversarial robustness, 3/
July 8, 2025 at 1:04 PM
The idea: instead of high-fidelity training from the get-go (the gold standard), we simulate the visual development from newborns to 25 years of age by synthesising decades of developmental vision research into an AI preprocessing pipeline (Developmental Visual Diet - DVD) 2/
July 8, 2025 at 1:04 PM
Second, we found these computations rely on just 0.5% of units. These units had learned to transform relative saccade targets into a world-centered reference frame. Lesioning them collapsed predictive remapping entirely. Instead, the model predicted the current fixation to also be the next. 5/6
June 5, 2025 at 1:14 PM
We make two important observations on emergent properties: First, the RNNs spontaneously learn to predict and inhibit upcoming fixation content. Energy minimisation alone drove sophisticated predictive computations supporting visual stability. 4/6
June 5, 2025 at 1:14 PM
Is this capacity genetically hardwired, or can it emerge from simpler principles? To find out, we trained RNNs on human-like fixation sequences (image patches and efference copies) on natural scenes. Only constraint: minimise energy consumption (unit preactivation). 3/6
June 5, 2025 at 1:14 PM
Can seemingly complex multi-area computations in the brain emerge from the need for energy efficient computation? In our new preprint on predictive remapping in active vision, we report on such a case.

Let us take you for a spin. 1/6 www.biorxiv.org/content/10.1...
June 5, 2025 at 1:14 PM
We have something in common then 😊

Among others, I keep telling people a specific story from this very book, suggesting that if the path is long, we need to focus on taking one step at a time and not be worried about how far in the distance the goal is, in order to be able to not panic and give up.
April 21, 2025 at 9:03 AM
#CCN2025 abstract acceptances were sent out this morning.

I'll post a summary of each of our projects closer to the conference.

Looking forward to seeing you all in Amsterdam!
April 21, 2025 at 8:40 AM
This year we ran another meme contest as part of my ML4CCN lecture series. Student submissions were fantastic.

Guess the paper...
February 6, 2025 at 7:26 PM
7/9 Let’s take this one step further. How does the neural response to saccade onset compare to stimulus onsets? Turns out that saccade elicited responses are **anticorrelated** to scene-onset responses, i.e. they produced opposite field patterns in source space.
November 1, 2024 at 4:12 PM
6/9 What we found was different. Instead of the fixation onset, the **saccade onset** explained most of the variance in latency and amplitude of the M100. That is, the bulk of the M100 (P100 also, as shown in a separate dataset) is NOT fixation-elicited.
November 1, 2024 at 4:12 PM
5/9 One of the most prominent, heavily studied components in visual neuroscience is the P100/M100, a prime candidate to compare active to passive vision. We expected this component to occur in response to fixation onset.
November 1, 2024 at 4:11 PM
4/9 To investigate neural responses of active vision, we went all in: head-stabilized MEG+high-res eye tracking, recorded while participants were actively exploring thousands of natural scenes.

This way, we collected more than 210k fixation events at high temporal resolution.
November 1, 2024 at 4:11 PM
3/9 For example, it is typically assumed that neural processes elicited by stimulus onsets are similar to fixation onsets.

Natural vision is different, however. Eye-movements are initiated by the brain itself, and where we look is less surprising than a random image stream.
November 1, 2024 at 4:10 PM