Multi-voxel pattern analysis (MVPA) is an increasingly popular approach for analyzing the results of fMRI scanning experiments that measure brain activity. MVPA searches for patterns of activation that correlate with a particular mental state. This is called ‘decoding’ neural activity.
Now a new paper in the Journal of Neuroscience from Caltech neuroscientists Julien Dubois et al. reports that MVPA is unable to decode certain kinds of information, even though single-unit recordings confirm that the information is present in brain activity.
In their study, Dubois et al. showed macaque monkeys images of male human faces. The images showed five men (five different identities) from five different viewpoints (angles). Here’s the stimuli:
In four monkeys, electrodes were implanted into the temporal lobe to provide single-unit recordings of neuronal firing. These electrodes were targeted at the face-selective region which is known to be involved in processing faces. Five monkeys got fMRI MVPA during the same task (but of these, only one monkey got both recordings.)
Dubois et al. found that fMRI MVPA was able to decode viewpoint information from brain activity with high accuracy. That is, based on the fMRI data, it was possible to tell whether monkeys were looking at a face that was frontal, or 45 degrees to the left, or 45 degrees to the right, etc.
However, MVPA could not decode the identity of the stimulus face. Yet the single-unit recordings did reveal the existence of neurons that selectively responded based on identity, as well as for viewpoint. In other words, face identity was encoded somewhere in the pattern of activity within the temporal lobe – but MVPA couldn’t decode it.
Why not? Dubois et al. consider various possible explanations for the failure of MVPA in the case of face identity. They conclude that differences in the spatial clustering of the viewpoint-selective and identity-selective neurons may be the answer.
Essentially, the theory is that the neurons that selectively respond to a particular viewpoint tend to be clustered together in a particular spot. The identity-selective neurons, however, are more scattered. Because fMRI has low spatial resolution, it may be unable to detect scattered signals.
By analogy, consider someone with poor eyesight. They might find it hard to spot a grain of sand, or even lots of grains of sand scattered over a surface, but they could see a pile of grains (a cluster).
Dubois et al. conclude that
Our study validated the notion that fMRI MVPA is a powerful tool for fMRI analysis. fMRI MVPA retrieved information about facial viewpoint with high fidelity.
However, we also unveiled a key limitation of fMRI MVPA in its failure to retrieve information about face identity [even though] this information was represented in the underlying neuronal populations…
Success of fMRI decoding depends strongly on the particular spatial organization of the variable being decoded.
What does this mean for MVPA more broadly? It suggests that the failure of MVPA to find a certain ‘code’ doesn’t mean that the code isn’t present. Absence of evidence is not evidence of absence. This is not really that surprising, but it’s an important reminder.
Dubois J, de Berker AO, & Tsao DY (2015). Single-Unit Recordings in the Macaque Face Patch System Reveal Limitations of fMRI MVPA. The Journal of Neuroscience, 35 (6), 2791-802 PMID: 25673866