Lorenzo Tausani

Our work "Stretching Beyond the Obvious: A Gradient-Free Framework to Unveil the Hidden Landscape of Visual Invariance" has been accepted at #ICLR2026 ! Thanks to P. Muratore @MorganBDTalbot @giacomo_amerio @gkreiman @davide_zoccolan for sharing this scientific journey ↓Thread↓

Our Stretch-and-Squeeze (SnS) paper has been accepted to #ICLR2026 @iclr_conf ! Credits to all the other authors for their great work: @LTausani, P. Muratore, @MorganBDTalbot, @giacomo_amerio and @gkreiman . Here’s a thread explaining SnS (full paper: arxiv.org/abs/2506.17040) 1/5

Stretch-and-Squeeze (SnS) has been accepted to @iclr_conf 2026! Our algorithm can be used by *both* neuroscientists and AI researchers to better understand *invariances* in neural networks: A thread about SnS:

5/5 By enabling the mapping of neuronal invariance without any prior information about the system, and also in a regime where the representation space is heavily subsampled, SnS sets the stage for a new way of studying invariance in neuroscience Paper: arxiv.org/abs/2506.17040…

4/5 We then tested how interpretable were the synthesised images in a human behavioural experiment. As it turns out invariances from robust models were classified more accurately under low-level stretching. However this advantage significantly decreased for high-level invariances

Our work "Stretching Beyond the Obvious: A Gradient-Free Framework to Unveil the Hidden Landscape of Visual Invariance" has been accepted at #ICLR2026 ! Thanks to P. Muratore @MorganBDTalbot @giacomo_amerio @gkreiman @davide_zoccolan for sharing this scientific journey ↓Thread↓

Very nice study from our lab. Out now on @PLOSCompBiol, check it out🧠👁👂

A new study from our lab, in collaboration with G. Kreiman @gkreiman group. We developed a feature visualization approach that allows exploring the selectivity and invariance fields of units in deep neural nets - with possible application to biological visual neurons.

1/ What if neurons could show you how they generalize? In this pre-print we introduce Stretch-and-Squeeze, where we let neurons map their own invariances instead of our biases L. Tausani @LTausani, P. Muratore, M. B. Talbot, G. Kreiman @gkreiman and D. Zoccolan @davide_zoccolan