A dead salmon perceiving humans can tell their emotional state. One mature Atlantic Salmon (Salmo salar) participated in the fMRI study. The salmon was approximately 18 inches long, weighed 3.8 lbs, and was not alive at the time of scanning.
Task. The task administered to the salmon involved completing an open-ended mentalizing task. The salmon was shown a series of photographs depicting human individuals in social situations with a specified emotional valence. The salmon was asked to determine what emotion the individual in the photo must have been experiencing.
Design. Stimuli were presented in a block design with each photo presented for 10 seconds followed by 12 seconds of rest. A total of 15 photos were displayed. Total scan time was 5.5 minutes.
And yet, so many people are incapable of the same thing.
Dead Salmon: 1, Humans: 0
That was my very first reaction.
So I guess this'll mark the beginning of people saying things like, "Aw c'mon, even a dead salmon could tell he was lying."
You are just finding articles to fit your songs now, eh?
Let this be a lesson: fMRI experiments are several layers of assumptions, one of which is a bunch of Bayesian math. Never trust a lone magnet experiment if theory doesn't back it up.
Unfortunately, fMRI generates pretty pictures and so mainstream media will run with a single magnet experiment's results over a dozen cognitive experimental results anyday.
It doesn't sound like they're taking a Bayesian approach since they talk in terms of p-values. Also, it seems like a prior putting close to zero probability on the dead fish thinking would be acceptable.
Whether or not there's a p-value has nothing to do with whether Bayesian analysis was used. The p-value is simply used to threshold the probability curve that's generated for each voxel. If Bayesian analysis is used, the probability map for each voxel just happens to be a Posterior Probability Map instead of a Statistical Probability Map.
For what it's worth, I went back and looked at this particular experiment. They used software that's capable of Bayesian but did not use it in this instance. They did use a p of 0.001 for each voxel, though, which in standard experimental procedures is considered extremely strong.
That is part of my point, though: fMRI involves *several* rather deep levels of assumption and analysis, *all* of which need to be handled correctly for the results to actually mean anything. Most of the time, most of the results are too far removed from the reality of the situation to actually be worthwhile. But they generate a pretty picture which everyone is more than willing to publish in mainstream media and say a single study proves everything.
1. BOLD assumes that changes in oxygen level are directly correlated with changes in blood flow. The magnet can only measure changes in oxygen level.
2. BOLD response is assumed to be representative of neuron activation. It is currently proven to correlate with blood flow, but it's not proven that blood flow is always correlated with neuronal activity.
3. BOLD response is based on relative signals, not absolute signals. Since it's based on blood flow, any activation can only be measured relative to prior activation levels. Therefore fMRI experiments all require a "baseline" activation map, which can be rather tricky to set up properly (and what if the baseline involves changes in blood flow that would cause negative flow instead of neutral flow in certain areas? You'd then miss activation there even if blood flow always correlates with activation!).
4. BOLD response is time-delayed from activation. After the neuron activates, the blood flow does not change for 1-2 seconds. The amount of time it takes is not mechanical and follows a curve, not an explicit digital signal. Therefore the "activation level" recorded will be different at slightly different times and it's difficult know for certain that it was in response to whatever stimulus was used 1-2 seconds ago.
5. fMRI images of BOLD responses are clustered into voxels. Now, 2-4 mm^2 voxels are pretty high resolution, but they're still nowhere close to the size of a neuron.
6. Statistical analysis is used to get around the fact that fMRI does not exactly scan every single voxel of the brain at the exact same time--they "realign the timeseries".
7. Normal MRI scans are used to figure out the anatomical layout of the brain. Statistical analysis of varying types is then used to line up the scanned fMRI voxels with anatomical layout from MRI.
8. Another set of statistical analysis is used to get around the fact that voxels are imprecise. In the above study they basically used Gaussian smoothing.
9. Another set of statistical analysis is used to generate probability models for each voxel, this can be Bayesian, but in the above was a linear model. This usually involves another round of "temporal readjustment" to, again, theoretically make up for all the timeshifting involved.
10. The probability models for each voxel are finally put through some more "normal" experimental math (t-constrast or similar), resulting in more "normal" probability curves for each voxel.
11. Those probability curves are then cut off based on hopefully appropriately-selected p-values.
Papers have been published arguing that there might be incorrect assumptions here.
For every single assumption outlined above.
Don't get me wrong, fMRI is really awesome, and I fully support its use. I just believe that the images generated are deceptively simple. Mainstream people see pictures and can casually inspect them, leading to a false sense of connection with brain activity.
Dead salmon is silently judging you.
The obvious implication is that Google should go into the salmon farming business and create a computing cloud of recently deceased salmon in order to better improve it's behaviorial tracking results.
You can probably sell adds 5% more effectively if you use dead fish computations as part of ad placement.
Perhaps, ultimately, each of us will have an individual strain of privately owned and operated dead fish which are selectively bred for our particular person.
In the end game, our human selves are simply proxies - avatars, if you will - for dead fish.
-t
They call this the Salmongularity.
You win.
'es only resting.
Excellent article title. Well done.
"By complete, random chance, we found some voxels that were significant that just happened to be in the fish's brain," Bennett said. "And if I were a ridiculous researcher, I'd say, `A dead salmon perceiving humans can tell their emotional state.'"
I will admit that this article cuts a little too close to the truth. I've seen academic Science in action. I've seen people squeeze their data in questionable ways to get Results. This isn't the first "Um, hey you guys?" study of this sort. Scientists need to be reminded of this occasionally. I sent it to a few friends in the fMRI lab that adjoins our EEG lab. Those poor fuckers and their semi-meaningless glowy voxels. Our squigly lines could beat them any day.