iiinova — where your digital self keeps evolving

We are incredibly excited about this because DVD may offer a resource-efficient path towards safer, more human-like AI vision — and suggests that biology, neuroscience, and psychology have much to offer in guiding the next generation of artificial intelligence. #NeuroAI @natmachintell.nature.com/fin

040

In summary, DVD-training yields models that rely on a fundamentally different feature set, shifting from distributed local textures to integrative, shape-based features as the foundation for their decisions. 10/

Result 5: In contrast to the previous literature’s main focus on visual acuity, we found that the development of contrast sensitivity is a key driver of shape bias in our models. 9/

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/

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/

Result 2: DVD-training enabled abstract shape recognition in cases where large vision-language models, despite being explicitly prompted, fail spectacularly. t-SNE nicely visualises the fundamentally different approach of DVD-trained models. 6/

Feature attribution using Grad-CAM 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/

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/

We then test the resulting DNNs across a range of conditions, each selected because they are challenging to AI: (i) shape-texture bias, (ii) adversarial robustness, (iii) robustness to image perturbations, and (iv) recognising abstract shapes embedded in complex backgrounds. 3/

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/

130

Daniel Anthes @anthesdaniel.bsky.social · 11/05/2026

Now out in Nature Machine Intelligence @NatMachIntell “Adopting a human developmental visual diet yields robust and shape-based AI vision”: doi.org/10.1038/s422.... A wonderful case where brain inspiration improved AI. With @martisamuser.bsky.social, Radek Cichy and @timkietzmann.bsky.social .

15222

Reposted by Zejin Lu

Excited about our new preprint: “The illusory simplicity of the feedforward pass: evidence for the dynamical nature of stimulus encoding along the primate ventral stream” arxiv.org/abs/2604.12825 Work with Sushrut Thorat, Anna Mitola, Paolo Papale, Peter König & Tim Kietzmann 🧵 thread below

arxiv.org

The illusory simplicity of the feedforward pass: evidence for the dynamical nature of stimulus encoding along the primate ventral stream

In studying primate vision, a large body of work focuses on the first feedforward sweep. During this initial time window, information is thought to pass through ventral stream regions in a stage-like ...

37524

Reposted by Zejin Lu

Victoria Bosch @initself.bsky.social · 03/11/2025

Introducing CorText: a framework that fuses brain data directly into a large language model, allowing for interactive neural readout using natural language. tl;dr: you can now chat with a brain scan 🧠💬 1/n

714253

Zejin Lu @zejinlu.bsky.social · 10/08/2025

If you are interested in development and development-inspired NeuroAI, and are coming to CCN this year, come join the workshop with us on 1️⃣1️⃣ Monday, Aug 11 🕒 3:00 – 6:00 pm 📍 Room A2.11 Register here: sites.google.com/view/child2m... (You can also come by my poster to chat!)

sites.google.com

CCN 2025 Satellite Event

Background The human visual system is full of optimisations—mechanisms designed to extract the most useful information from a constant stream of incoming data. The field of neuro-AI has made significa...

080

Reposted by Zejin Lu

Rhodri Cusack @rhodricusack.bsky.social · 25/06/2025

Not just one, but two fantastic chances to discuss how infant development can inform machine learning and vice-versa at CCN 2025 in Amsterdam!!! Satellite workshop sites.google.com/view/child2m... and Generative Adversarial Collaboration sites.google.com/ccneuro.org/...

sites.google.com

CCN 2025 Satellite Event

03113

Zejin Lu @zejinlu.bsky.social · 10/07/2025

Hi Lukas, very interesting work! Is it possible to know the shape bias in the Geirhos way? He reports the average shape bias across categories (see his plot code here: github.com/bethgelab/mo...). It would be even better if we could also know the average shape bias of each model across seeds:)!

github.com

model-vs-human/modelvshuman/plotting/plot.py at master · bethgelab/model-vs-human

Benchmark your model on out-of-distribution datasets with carefully collected human comparison data (NeurIPS 2021 Oral) - bethgelab/model-vs-human

010

Zejin Lu @zejinlu.bsky.social · 08/07/2025

🚨 Preprint alert! Excited to share my second PhD project: “Adopting a human developmental visual diet yields robust, shape-based AI vision” -- a nice case showing that biology, neuroscience, and psychology can still help AI :)! arxiv.org/abs/2507.03168

arxiv.org

Adopting a human developmental visual diet yields robust, shape-based AI vision

Despite years of research and the dramatic scaling of artificial intelligence (AI) systems, a striking misalignment between artificial and human vision persists. Contrary to humans, AI heavily relies ...

0133

In conclusion, All-TNNs are an exciting new class of networks for modelling primate vision, which address questions that are beyond the scope of CNNs and their topographic derivatives. 12/12

Next, we will use All-TNNs to explore the various factors of how smooth maps emerge from model training, without the need for a secondary smoothness loss. Possible avenues include wiring length optimization, energy constraints, local inhibition, or top-down connectivity patterns. 11/12

Can TNNs expand to self-supervised objectives? Yes, to a degree. We show that training All-TNNs with SimCLR yields smooth topography and category-independent spatial biases. However, SimCLR training fails to reproduce the structure of human-like category-specific spatial biases. 10/12

We show that these behavioural accuracy maps are structured and exhibit category-specific effects. Importantly, All-TNNs better reproduce these spatial structures of human visual biases than CNNs and other control models. 9/12

To study the impact of topography on behaviour, we conducted a human psychophysical experiment to quantify object recognition performance across spatial locations. This provided us with category-specific spatial accuracy maps for humans. 8/12

Similarly, All-TNNs allocate energy expenditure to task-relevant input regions, using an order of magnitude less “metabolic” cost than CNNs! And the smoother the topography, the greater the energy efficiency of the network! Energy efficiency was not explicitly optimised for. 7/12

Interestingly, All-TNNs exhibit a form of foveation, and allocate more processing resources to spatial regions rich in task-relevant information. 6/12

Upon training, topographical features reminiscent of the ventral stream emerge in All-TNNs, including smooth orientation selectivity maps in the first layer, and category-based selectivity clusters for tools, scenes, and faces in the last layer. 6/12

All-TNNs overcome this limitation. In All-TNNs 1) each unit has its own local RF, 2) units in each layer are arranged on a 2D “cortical sheet” without weight sharing, and 3) feature selectivity varies smoothly across space by encouraging similar selectivity in neighboring units. 5/12

Overall network architecture of All-TNNs

Yet, their reliance on weight sharing, i.e., detecting identical features across visual space, renders them unable to model central aspects of biological vision, such as the origin of topography and its relation to behaviour. 4/12

Background: CNNs are commonly used to model primate vision, and have been successful at predicting neural activity and at accounting for complex visual behaviour. 3/12

With Adrien Doerig(@adriendoerig.bsky.social) , Victoria Bosch (@initself.bsky.social), Daniel Kaiser (@dkaiserlab.bsky.social), Radoslaw Martin Cichy and Tim C Kietzmann (@timkietzmann.bsky.social). 2/12

In this work, we introduce All-Topographic Neural Networks (All-TNNs)—ANNs that drop weight sharing and learn on a smooth “cortical sheet,” capturing both human-like neural topography and visual biases in behaviour. 2/12

130