brainmapper

As a computational neuroscientist who has used ANNs to model brain data for 20 years, I want to suggest another way to think about these sorts of brain-ML/ANN comparisons: successful language comprehension requires representing the underlying statistical structure of language. Humans who learn to comprehend language must learn about the underlying statistical structure of language. ANNs that learn to comprehend language must learn the underlying statistical structure of language. Comparing human data to ANNs only really makes sense in the context of that statistical structure. What aspects of this statistical structure are learned by both machines (i.e., brains and ANNs)? What aspects of this structure are learned differently by the two machines? When the two machines learn different aspects of statistical structure, is this due to differences in architecture, training, or something else? A good understanding of the stimulus statistics is essential for interpretation.

Yes!!!

1) Before the 19th century science was the domain of the landed gentry or those who could find sponsorship from someone of means. Things have been different since then, but positions have always depended on academic pedigree. It has always been (and likely will continue to be) easier to get a permanent position at a great science University if your academic training ran through a great science University. This makes sense in some ways, but it also reflects residual bias (i.e. the halo effect of great Universities).

2) If anything, the system of recruiting for faculty positions has become MORE objective and meritocratic over the past few decades. Here is a true story that one of my older colleagues tells about getting his first academic job in 1964: a senior Professor called my colleague's Harvard PhD advisor and said, "We need someone in your area. Do you have any good students at the moment?" Six months later my senior colleague had a job at UC. That can't really happen any more (though the system is still far from perfect).

3) One thing that has changed is the ratio of applicants to positions. This has gotten a lot worse over the years. As a result, training has stretched out quite a bit. Postdocs are common/expected in many areas of science now, and they are often fairly long. But many of the most capable and persistent manage to find a way.

4) Always remember, science is ultimately about people. Therefore, it is inevitably political. Heck, if it weren't for the fact that, in the long run, science ultimately has to explain and predict things in the real world, it would be just as dysfunctional as Congress...

Oh my this article is painful to read on so many levels. It treats as important many phenomena that are either inevitable or uninteresting.

(1) The fact that engaging in one task distracts attention from another task is obvious. People derive inspiration from taking a shower, swimming, biking, playing music and yes, walking. All these tasks have something in common, they are different from whatever task one was doing beforehand, and they require less cognitive focus than many of the things that we do for work.

(2) The fact that noise in the brain is 1/f is not particularly interesting. Many natural systems have 1/f amplitude spectra. This pattern occurs commonly in multi-scale systems (the brain being one example).

(3) The fact that many aspects of brain activity (both signal and noise) are affected by aging, consciousness, mental experiences, memory and so on is not particularly interesting. Assuming that one is not a dualist, every distinct mental state must be reflected by some unique pattern of brain activity.

(4) Scientists know quite a bit about where these “spontaneous fluctuations” come from. Many of them are a consequence of changes in blood pressure, which varies substantially and randomly over time (within some finite band, of course). Others are caused by mental states that are difficult to measure and model, and so are unknown to the experimenter. But just because something is noise from the perspective of the experimenter doesn’t mean that it is noise from the perspective of the brain.

It wasn't designed to be the best one, it was just the first one. So every great scientist who wanted to live in a nice area with nice weather and great state support (which UCB had in the old days) went to UCB. Then when the other UCs were established it took a long time for them to work their way up. (Yes UCLA is also great! And UCSD and Davis are really good too.)

Yes! In 1980 the state of California paid about 80% of the cost of an undergraduate student's education. Today it is about 8%. The state has dis-invested from higher education to fund other things (like prisons).

As far as I can tell the UCB administration is not happy about increasing enrollment. But they don't really have control over this. The UC system decides how many students each campus must accept, and the campuses have to go along with it. UC Berkeley is not in control of this. (I work at UCB, and no one that I know is happy about the overcrowding.)

Sorry I should have clarified this comment was addressed at the claims in the article that most of the brain is activated even for trivial tasks...

Please note that merely seeing some place in the brain activate in a functional MRI task does NOT necessarily mean that that location is either necessary, sufficient or even involved in representing information relevant to that task. Functional MRI amplifies small global signals related to arousal, and if arousal changes during a task then these arousal-related signals can propagate over much of the brain. And even something as simple as an eye movement can be correlated with global changes in arousal. (A similar problem occurs with attention.) Unfortunately many of the most common modeling and analysis methods methods used in fMRI have no way to distinguish these rather uninteresting arousal-related changes from those that are actually informative about task-specific processes. The bottom line is that whenever you read about any fMRI result, you should ask yourself whether that could be a mere artifact of changes in arousal (or attention), and if so you should find out what was done to address this potential confound.

To quantify the value of a method it is useful to consider the amount of information that the method recovers from the data stream, the prediction accuracy of the resulting models, generalization ability outside of the conditions used to fit the model and decoding/reconstruction accuracy. For all four of those criteria, the best approach is to use linearized encoding models that estimate an FIR filter for each voxel separately.

As an fMRI practitioner, I just want to point out that this problem has nothing to do with fMRI. It has to do with the weaknesses inherent in the most common methods for designing fMRI experiments, and for analyzing and modeling the data. The SNR in fMRI depends heavily on the level of blood pressure, arousal, attention and other factors that vary day-to-day. Any analysis that doesn't account for that variability will be unreliable.

Methods for designing experiments, analyzing and modeling fMRI data that are robust to this variability are available. The problem is that most people in the field don't use them.

The most sensitive fMRI methods don't do alignment at all, they do all data processing in the individual subject's brain space.

Much of the day-to-day variation concerns uncontrolled variability in blood pressure, arousal and attention. Those factors have enormous influence on the blood-oxygen-level-dependent signal.

The movie reconstruction from brain activity work was from my lab, thanks for the shout out!

The problem with many methods of analyzing and modeling functional MRI data is that the signal-to-noise varies hugely across time, across individuals and across brain regions within an individual. Unfortunately, the most common methods of analyzing and modeling fMRI data do not incorporate any principled method for dealing with this SNR variability.

What makes this frustrating from a practitioner's point of view is that we do have methods for analyzing and modeling these data that can account for these uncontrolled SNR changes. The problem is that most people don't use these methods. (My lab has pioneered many of these techniques, and that is why we could produce those compelling decoding results.)