Ritchie, Kaplan & Klein (2019) 70 BJPS 581-607
The subtitle of this paper is “Neural Representation and the Limits of Multivariate Pattern Analysis in Cognitive Neuroscience”, but I read it anyway. It’s actually a very friendly paper which they could have called “MPVA decodability doesn’t entail explicit representation”.
The background is what they call the “decoder’s dictum”: that MVPA decoding of neural activity tells you what information the decoded patterns represent. The paper argues that this is false, for two basic reasons.
One is that MVPA decoding can pick up inessential correlations between the activity and the represented information. Those correlations might be in the stimulus—you can tell whether the subject is looking at up-right or up-left orientated lines, but perhaps you are picking up the output of edge detectors and things will fall apart if you make your stimulus patches square rather than round. Or they might be introduced by biases in the measurement procedure etc.
The second is that recoverability of information doesn’t mean it’s represented in a usable form. Visual input, for example, is fully informationally encoded in early processing (V1)—if it wasn’t, later stages would have to make it up. But it’s not usable in that form, it’s merely latent. Processing it into functionally available form is what later stages do.
The authors note that correlation with behaviour helps nail down usable content, but even that isn’t perfectly reliable. They suggest some furhter approaches which seem plausible enough to me. These involve connecting the neurological analysis with psychologically derived similarity relationships among the stimuli, or in the other direction with behaviour. This gladdens my pluralist heart; it doesn’t allow you to assign content based on neuroscientific data alone, but that is as it should be.