Purchase this article with an account.
Yasmine El-Shamayleh, Anitha Pasupathy; Size-invariant shape coding in visual area V4. Invest. Ophthalmol. Vis. Sci. 2013;54(15):3413.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
How do we recognize objects across changes in retinal size? This fundamental capacity of biological visual systems is computationally challenging. Size-invariant object recognition is supported by neurons in IT cortex, which maintain their preferences for objects across changes in scale. However, the detailed mechanisms that establish invariance remain unclear. To investigate this at the single neuron level, we targeted cortical area V4, a critical stage of object processing and the foundation of IT responses. Importantly, we leverage our understanding of object representation in V4 and a candidate model of shape encoding that makes direct predictions for neuronal responses at different sizes.
Many V4 neurons encode objects in terms of their component contour features; their selectivity can be modeled as preferences for contour curvature (convex/concave) at specific locations relative to object center (Pasupathy and Connor, 2001). This model presumes that neurons encode absolute curvature, a variable that is inversely related to object size; e.g., the curvature of a circle halves as its radius doubles. Thus, a curvature-tuned neuron cannot be size-invariant. This is because a particular contour feature will have different curvatures at different scales. We exploit this key idea here to ask whether neurons are curvature-tuned or size-invariant. We characterized well-isolated V4 neurons in two awake-fixating primates using parametric shapes that sampled a range of curvature values. Stimuli were presented at 4-5 scales and at 8 rotations, all inside the neuron’s receptive field.
We found that most V4 neurons were size-invariant, maintaining their shape preferences across the range of scales sampled (~2 octaves). Only a few neurons were curvature-tuned, shifting their shape preferences systematically, as predicted by the model, and reflecting selectivity for absolute curvature.
Our results motivate a key refinement of the curvature-based model of shape coding in V4; neurons are more accurately modeled as encoding contour characteristics relative to object size. This shape code may underlie the size-invariant object representation observed in IT cortex, and is well-suited to support size-invariant object recognition.
This PDF is available to Subscribers Only