Summary: meeting 2016 01 22 (Doug, Marlene) - mobeets/ContrastAttentionRFs GitHub Wiki

This morning I met with Marlene and Doug to talk about an email I'd sent them earlier, about somehow finding an appropriate way to publish the small bit of insight I gained during my rotation: Reverse correlation with V4 cells, or at least our V4 cells, is not that easy. In particular, after trying 5 or 6 different types of stimuli with full-field noise (think TV static, with binary intensity values), none of them revealed any receptive fields (RFs) larger than a few pixels using linear regression. But when we used sparse noise, i.e., only turning on one pixel in any given frame, the RFs were suddenly huge and beautiful. Here's an example of a few RFs, first using full field noise, then for sparse noise:

Now, whether the lack of interpretable RFs using full field noise is due to nonlinearities in the cells' responses, like surround suppression, we don't know. The V4 array is pretty foveal so it could just be that the true RFs are very tiny.

When we first set out to pilot these stimuli, we couldn't find any results on reverse correlation in V4. In fact, all we found was one tiny little paper, but this was just a proof-of-concept and not actual data. There was an SfN abstract promising results, but no accompanying paper.

So my thinking was, whether or not we know why our stimuli didn't work, any data-point we could put out there into the world could maybe help "the next poor shmuck" (Doug's words) who tries to do what we did. Even if we don't know why the data-point is where it is, or what the error bars were.

  • parallel to the software world, where if I'm trying to do something and all I come across is some abandoned Github repo where someone else clearly tried to do what I'm doing, even that repo alone can be really helpful.
  • What are the modes of publishing in science? Papers, talks, and posters. If it's not one of these three, it's nothing. And all of these three require lots of time. Quality over quantity.
  • the need for a data cemetery: somewhere I can just throw stuff up, to at least leave a landmark

Marlene's thinking was pretty different. It was a bigger picture, long-term kind of approach:

  1. Time is precious.
  2. All projects take time, even if you don't think they will, so you might as well pick something interesting to spend that time on.
  3. Is this really what you want to spend your time on?

But I was just looking for the lowest-effort way of putting something out there: Write up my stimulus parameters, some examples, some example RFs, and call it a day. I wouldn't want to spent more than a day on it. This tiny amount of effort seems reasonable to me because a) it will be nice to have a landmark of what I learned for when I come back to this project, and b) if I don't come back to this project, at least I offered something to the world! It's just unfortunate that there isn't an obvious place to share such a landmark. Writing up an SfN abstract, for example, would be the lowest-bar thing I could do to share this, but that clearly takes a lot of work. Instead, I might as well spend a bit of time exploring some more low-hanging fruit, like this one project Marlene suggested:

How are stimulus signal and noise represented differently in neural state space? I have repeats of each stimulus I showed to the V4 array. So for each pair of stimuli, calculate the vector connecting the spike count responses to each of those stimuli. If the pair of stimuli were identical, this vector is a noise vector. If the pair of stimuli are different (and we could think of a continuous description of "how different", e.g., using their spectral properties), this vector is a signal vector.

So how do these distributions of vectors differ, for signal vs. noise? Are their dimensionalities different? Or, even easier: Are their norms different?

Just a bit of poking around could maybe lead towards a way more interesting SfN abstract than "Hey, we tried reverse correlation with full-field noise in V4 and it didn't work!"