Hue tuning curves in V4 change with visual context¶
Foreword
A common method of investigating visual cortex is to characterize whether neurons are ‘tuned’ to visual features. How much does this approach tell us a neuron’s general role in vision?
Tuning curves allow generalizing from stimuli by making assumptions about responses to unseen stimuli. One key assumption is that the encoded parameter (e.g. orientation) affects activity in a similar way on many stimuli. One way to verify such a claim is to measure tuning in other contexts. If tuning describes a general role in processing, it should generalize to stimuli not presented.
This chapter provides a method to establish how far a tuning curve might generalize. It is an example of Role 1: machine learning as an engineering tool for neuroscience.
One way to confirm this assumption is to measure tuning in other contexts. In vision, the context of natural scenes is the ethologically relevant one, but it is difficult to measure tuning on natural scenes because each image is different in many ways (not just the tuning curve’s feature). The nonlinearity of the cortical response also complicates this effort. Here, we developed a method to estimate tuning despite these difficulties. This can help to establish how far a tuning curve might generalize.
This chapter was presented as a poster at the Bernstein Conference for Computational Neuroscience in 2019 and the Conference of Cognitive Computational Neuroscience in 2017.