Category Archives: Journal club

Category representation in the brain

Posted on by

figure3

I was at a web development conference several years ago and went to hear the keynote address by Rasmus Lerdorf, the developer of PHP. He was working for Yahoo! at the time, and told us this succulent anecdote:

Some new guy came around and asked what kind of database we used to store the categories so people could search them. So I answered him:

- Oh, there’s no database, it’s a filesystem. The categories are folders. When users search through them, the server runs calls grep in the background.

He looks at me, horrified.

- Grep?
- Well it’s a fast filesystem, and it works.
- What if you want more features?
- Don’t need ‘em.
- But, it must be slow. What do you do if you want to make it faster?
- Buy more computers!

Indeed, you can solve most problems in life by buying more computers – or grad students, or microscopes, or lasers for that matter. Some of Jack Gallant’s lab recent efforts in fMRI analysis are a good example of this approach. They published an interesting paper in Neuron last month about the representation of categories in cortex. Continue reading →

Fitting a convolutional model of V1

Posted on by

Fresh off the press from NIPS 2012 is a paper by Vintch et al on a convolutional model of V1. Unsurprisingly, the convolutional model (or as they call it, the subunit model) beats the crap out of an STC-based model since it captures the same basic features of the RFs in a much lower dimensional fashion than the STC model. Indeed a model of this class by yours truly has been leading the Neural Prediction Challenge for a couple years.

The main issue with convolutional models is that they’re a bitch to fit. Vintch et al. solve the problem by using what they call convolutional spike-triggered covariance. The idea is to form an augmented stimulus matrix which contains all the values within each possible window underlying the convolution. The augmented stimulus matrix is itself point-wise multiplied by a guessed spatial kernel corresponding to the receptive field envelope. Then STC is run on that, and the lowest and highest eigenvalue-associated eigenvectors are used as initialization for the filters of the convolutional mode – they use two kernels only for the convolutional net.

It’s an interesting technique, and it seems like it’s the right idea, but it’s really ad hoc. I would have liked to see it fleshed out, or at least see some simulations to understand under what circumstances it picks decent initial filters. I’m sure the usual suspects will take a stab at it.

I think convolutional models are the way to go to explain early and intermediate visual areas, but there’s a couple of unsolved issues:

  • How do you regularize the solutions? Granted, the models are lower dimensional than STC, but they do state that they have “only” 1200 parameters. I’ve been experimenting with a modified version of boosting for the parameters of the kernels, but it’s very computationally intensive — takes a day to fit 5 minutes of data. L1 seems like a pain as well.
  • How do you speed up the fitting process so that you can evaluate a lot of different model variants? The augmented data matrices are huge so fitting the kernels is very expensive. I was wondering whether it was worth it to keep the convolution matrix implicit and implement the matrix-vector products via convn or fftn. At the very least it would take less memory so that you could parallelize the inference.
  • What about a nonlinearity at the very start of the cascade? Is that important? By an appropriate choice of input and intermediate nonlinearity, you can get AND or OR-like integration. Been working on an NLNLN model of this type this summer, but have yet to analyse the results (*damn you NY/Montreal girls!*).

Anyways, happy fun times in RF estimation land.

Modifying body representation through vision and vice versa

Posted on by

There’s a few interesting multi-modal illusions involving vision and another sense. Proprioception, the ability to sense the position of one’s body parts, is one sense that gets a bad rep; it’s not even included in the classic 5 senses. Yet it’s certainly quite important, and people that lose proprioception have difficulty functioning at first, a condition that is mitigated over time through vision, as related by Oliver Sacks.

Can you modify vision through proprioception, and vice versa? There’s a neat illusion related in Carlston et al. 2010 that shows just this. A subject sits in the dark for several minutes. A blinding, 1 ms flash then occurs; after a certain number of seconds, the person perceives a positive afterimage of the scene. If the scene includes one’s own limbs, then interesting things happen: if you move the limb away from its original position, it seems to fade in the afterimage. If you move it back to the original spot, it seems to fade back in.

The fading appears to extend not only to one’s own limbs, but also to tools. You can hold a ball in your hand, for example, then drop it, and it seems to be removed from the afterimage. Admittedly this is hard to believe, so I tried to replicate the experiment for myself and see what it’s all about. I don’t have the 3,000$ flash system used in the original experiment, but I did buy a 20$ flash in a pawn shop. It took a little while to get the effect going, but basically the key is to hold the flash very close to the limb whose image is to be manipulated. Then a very clear blue-black afterimage appears after a few seconds, clear enough to try some of the effects described in the paper.

Here’s a video showing how to get it running:

It’s perhaps not as mindblowing as it appears in the paper; actually I don’t really like how the paper is written because there’s few actual experiments and the style is bombastic (it’s not very sciency). But the effect is real nonetheless, and I’m sure it would be more impressive with a 2000W lighting system handy.

The same setup can be used to create illusions of floating limbs, as related in this abstract by Ramachandran. It’s pretty cool, and easy to get going, you should try it out.

ResearchBlogging.org

Carlson, T., Alvarez, G., Wu, D., & Verstraten, F. (2010). Rapid Assimilation of External Objects Into the Body Schema Psychological Science, 21 (7), 1000-1005 DOI: 10.1177/0956797610371962

Hearing radio frequencies

Posted on by

I was reading the Wikipedia article on tinnitus, and came across this pearl of a sentence:

A common and often misdiagnosed condition that mimics tinnitus is Radio Frequency (RF) Hearing in which subjects have been tested and found to hear high-pitched transmission frequencies that sound similar to tinnitus.

Hmm, what? Yes, humans, under special circumstances, can hear radio-frequency pulses in the range of 2.4MHz to 10GHz (corresponding to radio frequencies and microwave) as buzzes, clocks, hiss or knocking at apparent auditory frequencies of 5kHz and higher (very high-pitched). That doesn’t mean that you can hear talk radio by receiving AM waves (that would be unbelievably annoying); it just means that when it’s very very quiet, you can hear a faint high-pitched noise from RF sources.

But how could electromagnetic waves be detected as sound, which is a pressure wave? After all, light is an EM wave too, but we don’t hear light! It’s a long story, but basically, you’re a microwave bongo head. Elder and Chou (2003) offer a thorough overview of the phenomenon.

Apparently, RF hearing was first reported in the 1940s by people working with radar, but reports were dismissed as illusions or hallucinations. The phenomenon was investigated scientifically by Frey in 1961, who concluded that RF hearing is a real thing. It can be stopped, for example, by placing a piece of aluminum between the RF source and the ear.

RF sources can only be heard by people with working audition above 5kHz. This would imply that RF sources create an acoustic vibration close to the cochlea that gets detected as high-frequency sound. Indeed, one can record electrical potentials inside the cochlea evoked by RF pulses that look just like potentials evoked by sound waves.

The authors further report that the apparent acoustic frequency of the RF pulse is independent of the EM frequency of the actual pulse but dependent upon head dimensions. So EM energy gets absorbed by the head and somehow this energy is transformed into pressure waves that get reshaped by the head. Thus, microwave bongo head.

The most likely explanation for this is the thermoelastic expansion theory. When RF pulses are created near a container of water, it is possible to detect evoked sound waves in the water; the acoustic frequency of these waves is similar to that of the sounds heard in RF hearing. When an RF pulse is absorbed by water, it locally elevates the temperature, which causes a rapid local expansion which then gets propagated as a pressure wave. The local elevation in temperature can be quite small: the authors give a figure of 5 x 10-6 degrees Celsius (!). This sound wave gets transmitted by bones to eventually make its way to the cochlea, where it gets detected as just another pressure wave.

The authors point out that this is neither dangerous nor useful. It’s just kinda cool. Ain’t science neat?

ResearchBlogging.org

Elder, J., & Chou, C. (2003). Auditory response to pulsed radiofrequency energy Bioelectromagnetics, 24 (S6) DOI: 10.1002/bem.10163

Modeling the Spatial Reach of LFP

Posted on by

As regular readers will know, I’ve been intrigued by the nature of local field potential for some time. There’s a recent paper in Neuron by Lindén et al. that uses a modeling approach to explain the spatial reach of the LFP. This is a subject of some controversy; the spatial reach of the LFP has been estimated at less than 250 microns by some and several millimeters by others. The paper resolves these discrepancies by showing how the apparent integration size depends heavily on these (and several other) factors:

  • correlations: if the underlying population is correlated, the LFP reaches further than if the underlying population is uncorrelated
  • position of the recording electrode: the greater the difference in depth between the electrode and the location of the synaptic input, the greater the apparent reach of the LFP

So in short, there is no such thing as THE spatial reach of the LFP; it depends on a bunch of things and over an order of magnitude. The paper is fantastically presented, and while I haven’t gone through the methods in details, it looks pretty legit.

If I may so speculate, the widely varying estimates in the literature are probably due to different experimental paradigms. For instance, the reverse-correlation-type experiment used in Nauhaus et al. 2009 might encourage a less correlated population, while the natural images and longer presentation times in Kreiman et al. 2006 may encourage a more correlated population.

It will be interesting to see how their findings are affected once they consider frequency-specific spatial reach, although thismay require making more than a few assumptions concerning the dynamics of the synaptic input. Anyhow, read the paper, it’s good.

ResearchBlogging.org

Lindén H, Tetzlaff T, Potjans TC, Pettersen KH, Grün S, Diesmann M, & Einevoll GT (2011). Modeling the spatial reach of the LFP. Neuron, 72 (5), 859-72 PMID: 22153380