Visual Perception and Neural Processing

Visual perception emerges from a cascade of neural transformations that begin with photoreceptors in the retina and proceed via the lateral geniculate nucleus to primary visual cortex (V1), thereafter diverging into ventral and dorsal pathways. The ventral stream underlies object recognition and form perception by extracting increasingly complex and invariant feature representations, while the dorsal stream processes motion, spatial relationships and action-guided behaviour. Neuronal populations across cortical areas employ feedforward sweeps, recurrent loops and lateral interactions to achieve constancy under changes in illumination, scale and viewpoint. Computational frameworks—ranging from efficient-coding principles to predictive-coding theories—seek to explain how perceptual stability and rapid categorisation arise from dynamic population activity. Insights from neuroimaging, electrophysiology and psychophysics have converged with advances in deep neural networks to illuminate the algorithms and circuit motifs that support human visual function and inspire artificial vision systems.

Research from Nature Portfolio

Recent studies have introduced the concept of model metamers—artificial stimuli optimised to match network activations at specific model stages—to probe divergences between machine and human invariances. These investigations reveal that, although late-stage deep network representations may appear analogous to cortical patterns, their metamers can be unrecognisable to human observers, highlighting idiosyncratic invariances and suggesting new benchmarks for evaluating sensory models. In parallel, work combining magnetoencephalography and functional MRI has demonstrated that layers of convolutional neural networks recapitulate the temporal and spatial progression of human visual processing: early network layers correspond to prompt responses in V1 and V2, while deeper layers align with later activity in inferior temporal and parietal regions. Such findings underscore the importance of ecologically valid training regimes to reproduce the brain’s hierarchical dynamics.

Visual Perception and Neural Processing publication trend

The graph below shows the total number of articles in visual perception and neural processing across all publications each year (not limited to Nature Index journals).

Technical terms

Model metamer: An artificial stimulus generated to elicit identical activations at a given layer of a neural network, used to compare machine and human perceptual invariances.

Deep neural network: A multilayered computational model inspired by cortical circuits, in which successive layers learn hierarchical feature representations from data.

Inferior temporal cortex (IT): A high-order region in the ventral visual stream critical for object recognition and categorical representation.

Ensemble perception: The ability of the visual system to extract summary statistical information (e.g. mean size, orientation) from groups of objects rapidly and efficiently.

Transfer learning: A technique wherein features learned by a network on one task (e.g. object categorisation) are applied to improve performance or modelling on a different task (e.g. neural response prediction).

References

1. Model metamers reveal divergent invariances between biological and artificial neural networks. Nature Neuroscience (2023).
2. Comparison of deep neural networks to spatio-temporal cortical dynamics of human visual object recognition reveals hierarchical correspondence. Scientific Reports (2016).
3. Ensemble Perception. Annual Review of Psychology (2017).
4. Deep convolutional models improve predictions of macaque V1 responses to natural images. PLOS Computational Biology (2019).