Spontaneous Emergence of Brain-Like Direction Maps in Artificial Neural Network
A new study from arXiv (2605.11718) demonstrates that direction-selective maps resembling those in the primate middle temporal (MT) area can emerge spontaneously in a spatiotemporal Topographic Deep Artificial Neural Network (TDANN). The researchers trained a 3D ResNet on naturalistic videos using a Momentum Contrast (MoCo) self-supervised paradigm combined with a biologically inspired spatial loss. The resulting network developed brain-like direction maps and topological pinwheel structures, revealing that MT topography may be governed by universal principles similar to those in the ventral stream. The work addresses a longstanding question in neuroscience about the computational origins of dorsal stream topographies. Key findings include strong direction selectivity paired with residual tuning properties, suggesting that spatiotemporal contrastive optimization is sufficient to replicate key features of MT organization. The study was published on arXiv under ID 2605.11718.
Key facts
- Study published on arXiv (ID 2605.11718) on an unspecified date.
- Research uses a spatiotemporal Topographic Deep Artificial Neural Network (TDANN).
- Model trained on naturalistic videos via Momentum Contrast (MoCo) self-supervised learning.
- Biologically inspired spatial loss incorporated during training.
- Network spontaneously developed direction-selective maps and pinwheel structures.
- Maps resemble those in primate middle temporal (MT) area.
- Suggests MT topography may follow universal principles like ventral stream.
- Key finding: strong direction selectivity with residual tuning properties.
Entities
Institutions
- arXiv