LCN controls, which have an identical architecture to All-TNNs, but lack the spatial similarity
loss do not develop topographic features that are similar to the brain, despite being trained
on the same dataset and categorization objective. This ties into an important current debate
about which aspects of the visual system's structure are genetically hardwired, and which
require experience 36,51 , and lends support to the idea that both visual expertise (here:
dataset) and the right inductive biases (here: architecture and spatial loss) are necessary
driving factors for the emergence of functional topographies in the brain. All-TNNs thus invite
further modelling of the visual cortex and beyond 52 , taking developmental genesis into
account through systematic manipulations of architectural features, loss functions, or training
datasets to uncover principles and mechanisms underlying the maturation of brain structure
and behaviour in the visual system 53 .
Importantly, the topographic features of All-TNNs are central to replicating important aspects
of human behaviour. Humans can exploit spatial regularities in the typical locations of
objects, which may be an adaptive strategy to reduce the computational load and enhance
visual efficiency in complex environments 26 . In line with this strategy, neural representations
are better decodable and perceptual sensitivity is higher when objects appear at locations
that match their typical positions in the world, allowing objects to be more easily detected
and recognized when presented in expected locations 26,54 . We find that All-TNNs are closer
to human visual behaviour in this setting, due to how their topographical organisation
impacts object recognition.
The ability of All-TNNs to link topographies to behaviour allows for new research directions.
An obvious hypothesis to test is that if certain objects have more importance than others
during training, they may take up more space in the topography, which in turn may account
for biases in behaviour 55–57 . As another example, the spatial topography of All-TNNs allows
for targeted lesioning of its organised parts, such as the face-selective units in the later
layers. This allows using All-TNNs to model brain lesions 58 with potential for clinical impact,
and a model of virtual lesioning methods such as Transcranial Magnetic Stimulation (TMS) 59 ,
providing insights into the underlying mechanisms and effects of such experimental
interventions.
All-TNNs complement other recent approaches to modelling topography in the visual
system 30,41,60–66 , which have greatly contributed to our understanding of cortical map
formation. One limitation of most existing models of topographic organisation is that they are
either truly topographic but not task-performing or task-performing but not truly topographic.
Examples of the former are hand-crafted self-organising maps 13,67,68 . The latter are most
often, if not always, based on augmenting CNNs, for example by adding a spatial remapping
to their units, building self-organising maps based on unit activities, or creating multiple CNN
streams 30,60,61,64–66 . While we strongly agree that CNNs can provide important insight into
functional organisation, such models do rely on biologically implausible weight sharing rather
than genuine topography. All-TNNs are a promising approach to overcome this limitation, as
they are both mechanistically truly topographic and task performing.
The current work has several limitations. First, the spatial similarity loss that we used to
encourage neighbouring units to learn similar features may not capture the mechanism of
smoothness in topographic organization and the exact spatial arrangement of topography in
the brain. Future work using All-TNNs as a starting point can explore how smooth maps can
emerge naturally from model training, without imposing a secondary spatial similarity
14