,
9 tweets,
7 min read
Read on Twitter
New preprint! Does V4’s tuning to hue on artificial stimuli tell us about its tuning to hue on natural images? Not really. (Twitterified in the thread below) biorxiv.org/content/10.110… with @desipoika, @jugoguj, Matthew Smith, and @KordingLab 
@desipoika @jugoguj @KordingLab So, a classic way to ask "what does neural activity mean?" in the visual cortex is to vary stimuli in a systematic, parameterized way while recording responses. But do those results tell you what activity 'means' in other contexts, like while viewing natural scenes?
@desipoika @jugoguj @KordingLab In V1 there’s a rich literature asking this question by estimating tuning from natural scene responses. But what about in higher areas, like V4? We can’t use the same methods, since responses are so nonlinear.
@desipoika @jugoguj @KordingLab It took a few years to get it right, but we managed to estimate tuning of macaque V4 neurons from natural scenes. We looked at hue tuning, since there's a already big literature about color responses in V4.
@desipoika @jugoguj @KordingLab We tried reverse-correlation approaches and also a new method: we fit a nonlinear V4 model (based on ImageNet-trained DNNs, Yamins/DiCarlo style) and characterized how tiny perturbations to hue affected the model’s response.
@desipoika @jugoguj @KordingLab We then compared these tuning curves to hue tuning curves estimated using artificial stimuli. Were they similar? Not really. Stated cleanly: in different visual contexts, the immediate effect of varying hue upon the response was different.
@desipoika @jugoguj @KordingLab This is obviously about how we understand V4’s activity. But we also think this is a cautionary tale for the vary-stimuli-in-a-parameterized-way approach. When does this approach tell you about a neuron's response on other stimuli & the neuron's general role?
@desipoika @jugoguj @KordingLab It’s a hard question that we’ve thought carefully about. There’s an empirical and a theoretical side. Our thoughts are in the Discussion, but please, tell us your thoughts!
Ah Matt I couldn't find your handle! Everyone should follow @SmithLabNeuro
