I’m introducing a new (hopefully recurring) feature on the blog: the Journal Club express. Lengthy discussions of papers are quite time-consuming to write, so instead I’ll periodically highlight a few recent papers I’ve read that I think could be interesting to regular readers.

There’s a new Neuron paper from John Maunsell’s lab that shows an intriguing link between attentional modulation and tuned normalization. They recorded in MT while the subject did a direction change-detection task. By manipulating the locus of attention and the contrast and direction of targets both within and outside the receptive field, they found a strong link between attentional modulation and the strength of normalization within these receptive fields, consistent with earlier findings. Furthermore, they show that the strength of attentional modulation is best predicted by the extent of tuned normalization within the receptive fields.

In their framework, tuned normalization differs from regular normalization in that the strength of normalization is different for stimuli going in the null or preferred directions. The results are quite convincing, but I don’t really like how they constantly reference Rust et al. 2006. The specific form of tuned normalization used in Rust et al. 2006 differs from that in the new Maunsell paper in that it is 1) completely equivalent to a static compressive nonlinearity and 2) affects V1 afferents and not MT cells themselves. I understand that there’s a link but it seems to me it’s less strong than they make it out to be.

Also in the stack of papers about to fall off my desk, the latest perspective on hierarchical object processing in Neuron by Jim DiCarlo, Davide Zoccolan and Nicole Rust. It can be viewed as a follow-up on DiCarlo and Cox (2007). The 2007 paper I think was kind of mind blowing, introducing the concept of invariance as a means of untangling high-dimensional object manifolds. Reading the original paper, I thought we had a new blueprint for studying high-level visual cortex. To me, the original paper screamed: to study object recognition, you must study both object manifolds in a theoretical way and response isosurfaces in high-level visual cortex. I wish that somebody would do the same sort of closed-loop measurements of isosurfaces in V4 and IT that Bölinger and Gollisch have recently done in the retina.

Somehow this aspect of research hasn’t panned out in the expected way. There’s some discussion of the latest results from Nicole’s stint at DiCarlo’s lab on the increase of tolerance and selectivity between V4 and IT in the new paper, some links between their perspective and slow feature analysis, and a more concrete instantiation of the model in terms of anatomy. I’m dumbfounded by their claim that looking at single neurons is (depending on how you interpret it) uninteresting or foolish.

Finally, there’s an awesome paper in Nature Neuroscience by Weber, Machens and Borst about optic flow selectivity in flies. This a very nice application of Generalized Linear Models in understanding both functional connectivity and stimulus selectivity in a population of 2 (!) neurons. There’s some great ideas here on how to use GLMs as part of an information theoretic analysis.

Papers mentioned: