Variable Density, série Delaunay | Next-generation neural computations

À la croisée des arts plastiques et de l'architecture, Etienne Rey explore les dimensions comme une matière première propice à la construction de formes évolutives. A travers ses œuvres où l'expérience est primordiale, Etienne Rey jongle avec les composantes de la lumière, de l'espace et pour certaines également du son. La démarche de ce sociétaire de l'Ososphère tend à révéler les mécanismes d'organisation, d'interrelation et d'interaction qui définissent notre monde.

Tales from the Poles - APECS Netherlands

Tales from the Poles Tales from the Poles is a collaborative storytelling project organized by APECS Netherlands and the UK Polar Network. Its aim is to bring together diverse voices connected to the Arctic and Antarctic, including scientists, Early Career Researchers (ECRs), community members, station staff, artists and anyone who has lived

Kaalstaart

Kaalstaart

A New Look for Convolutional Deep Networks | Next-generation neural computations

By construction, deep convolutional networks treat all pixels equally. Result: they are energy-hungry, fragile to rotations or zooms, and poor at object localization. This contrasts sharply with biological vision, which demonstrates remarkable efficiency. Think of an eagle spotting prey from kilometers away or a human recognizing a face within an eye blink time. We propose that by using a mechanism concentrating processing at the center of gaze, biological vision is optimally suited for object localization. To test this hypothesis, we embed foveated retinotopy into CNN inputs. We observe that while standard ResNets drop from 80% accuracy to 46% under rotation attacks, our retrained networks achieve comparable accuracy (76%) while dropping only to 66%. Critically, without additional learning, this foveated CNN reveals high sensitivity to fixation point location: the maximum likelihood position points to ground truth bounding boxes with 85% accuracy, compared to 71% for classical CNNs. This suggests that this specific Nature's million-year solution may offer deep learning visual networks a fundamentally new architectural look.