I want to react to a twitter post that tweeted a statement from a talk during the Cosyne meeting, to the effect that “large scale neural recordings are pointless unless behavior is complex. Else dimensionality is too low” (tweeted by @anne_churchland). I won’t name the speaker because there is nothing to gain by doing so, but I would ike to explain why that statement is totally misguided.
The value of large-scale recording is that it allows an experimenter to look at internal relationships. Instead of merely looking at the relationships between brain cells and external variables, you can look at the relationships of brain cells to each other.
In fact, if what you care about is the relationship between brain cells and external events, there is really nothing to gain from large-scale recordings other than mass production of data.
But if you are interested in internal relationships, large-scale recording has tremendous value. Even if you are only interested in the relationships between pairs of neurons, the number of pairs in a dataset is proportional to the square of the number of neurons. If you are interested in larger groups, the benefits are even greater.
This is not just a theoretical statement; it reflects my own practical experiences. I was a graduate student and postdoc in Bruce McNaughton’s lab during the period when “hyperdrive” recording was developed (that is, the technique of using 12 tetrodes at once to record simultaneously from populations of up to 100 or more rat hippocampal cells). I was able to acquire and analyze hundreds of data sets using that technique, in a wide variety of behavioral paradigms.
My experience was that the most useful data came from the simplest tasks — in fact the recipe for getting the most out of an experiment was to simplify the task as much as possible and maximize the amount of data collected as far as possible. The most useful data set I ever collected came from a rat running round and round a triangular track for an hour, stopping in the middle of each arm to eat a small food pellet. (See this paper for some of the results.)
With large-scale recording it is not necessary to have any task at all in order to gain valuable information. Even the brain of a sleeping animal shows extraordinary complexity when examined on a sufficiently large scale.
Using complicated tasks with large-scale recording is actually likely to be counterproductive. If the goal is to examine internal relationships, then complexity works against you. The more complicated the task, the more variability will be attributable to external factors rather than to internal relationships.
Bottom line: the essential error in the speaker’s statement is in the “else dimensionality is too low” part. Brain activity has very high dimensionality without requiring any external task to increase it. People who use fMRI to study the so-called “default mode network” have begun to realize it, but recording from large ensembles of individual neurons brings the message home even more strongly.