Wednesday, July 15, 2009

Discussion with Friedemann Pulvermuller

In response to a recent blog entry and subsequent discussion of a paper by Friedemann Pulvermuller, Friedemann himself has joined the discussion to clarify a few issues. His commentary is important enough to move it out of the "basement" so I have copied it here and provide my own response below. I hope we can continue this discussion in the comments...

From Pulvermuller:


You miss the point of our HBM paper. Our study for the first time shows using an obective data driven method that the semantic circuits, the semantic cell assemblies if you wish, are distributed - and even more distributed than many of us dared to think before.

The Hickok Poeppel model does not account for such distributed and category specific semantic cirucits. Maybe you should change it?

BASIC: If the voxel of *maximal activation* is slightly besides your prediction - what does this mean in the face of really huge activation clouds? You must show that the could DOES NOT REACH INTO the premotor and motor cortex. Look at David Kemmerer's recent plot: Indeed some find somatotopic semantic activation outside the predicted fields - but most studies show good corresponence between motor areas and action/body-semantics.

BTW, as previous contributors have emphasied, we have used motor localiser and the finger movement area. Hand word activations do indeed overlap. The mapping to standardized motor maps is indeed irrelevant here.


Hickok response:

Hi Friedemann,

Great to hear from you! I hope you will continue posting here as we will all benefit from clarification of theoretical positions, details of the data, etc.

First of all, contrary to what you claim, the Hickok-Poeppel model is not a model of conceptual semantics. Our theoretical coverage stops at the interface between acoustic-phonological systems and conceptual systems.

Second, despite the fact that our model doesn’t deal with conceptual semantics directly, we have consistently and repeatedly stated our assumption (without defense and deferring to people like you who do research in this area) that conceptual systems are widely distributed.

Here are some quotes from our papers (emphases added):

“This pathway appears to be important for interfacing sound-based representations of speech with widely distributed conceptual representations” (Hickok & Poeppel, 2000, p. 131)

“The multimodal cortical fields in the vicinity of the left temporal-parietal-occipital junction are a good candidate for networks important for interfacing sound-based representations in auditory cortex with widely distributed conceptual–semantic representations” (Hickok & Poeppel, 2000, p. 134)

“These pITL structures serve as an interface between sound-based representations of speech in STG and widely distributed conceptual representations” (Hickok & Poeppel, 2004, p. 72)

“…semantic information is represented in a highly distributed fashion throughout the cortex, and middle posterior temporal regions are involved in the mapping between phonological representations in the STS and widely distributed semantic representations” (Hickok & Poeppel, 2007, p. 398)

I don’t know how to state it any more clearly.

It puzzles me why you believe the Hickok-Poeppel model assumes a non-distributed conceptual system. In the same way, it puzzles me why you believe the classical neurologists such as Wernicke & Lichtheim assume non-distributed conceptual system (see previous blog entries). These models certainly don’t need to be changed to account for your data. Maybe you could explain where your misconceptions are coming from?

Now, quite aside from the Hickok-Poeppel model, let’s talk about how conceptual information is represented. As I’ve stated in print, I believe in widely distributed systems that include both sensory- and motor-related regions. My guess, however, is that these are higher-level systems and not, as you propose for action concepts, low level motor representations (e.g., in M1). My belief is based primarily on the many-to-many mappings between low-level action coding and action concepts, and I have not yet seen any experimental evidence convincing me that this believe is incorrect including your TMS, imaging, or lesion-based studies.

Greig’s comment on your recent paper is relevant in that your general claim is not just that conceptual representations are distributed (an old idea that is widely accepted) but that it involves low-levels of the motor system in a somatotopic fashion. To make such a claim for a given result one either has to show that the action concept processing activations are in motor cortex (and imprecise method as you point out) or show direct overlap between motor actions and conceptual processing. Your recent paper seems to have done neither (I admit I haven’t yet read it, so correct me if I’m wrong).

1 comment:

Greig de Zubicaray said...

I’m glad to see this discussion is gathering pace and I’m happy to respond to two of Friedemann’s points:

In response to the question “If the voxel of *maximal activation* is slightly besides your prediction - what does this mean in the face of really huge activation clouds?”, I’ll quote directly from Eickhoff et al. (2007) who addressed this question empirically:

“First, since the functional hypothesis is tested for each voxel separately, inference (which includes inference on the anatomical location) should be restricted to individual voxels. Moreover, due to the applied spatial smoothing, each voxel is a weighted mean of its local neighbourhood. Therefore, the local maximum represents already the region, where the assessed effect is most pronounced, eliminating the further need to examine other neighboring voxels. Finally, by using the local maximum as a representation of the underlying activation, one can circumvent the dilemma that different thresholds or thresholding strategies will result in differently sized activations: whereas the extent of activation is variable, its peak response is not. In other words, the size of the iceberg depends on the level it is viewed at, whereas its tip remains constant.” (p. 519)

Note that Postle et al. (2008) plotted relatively large 10 mm spheres around the peak maxima from several studies, corresponding to several voxels.

The next issue is what constitutes a valid map of motor cortex activation, a cytoarchitectonic one or one derived from functional neuroimaging studies (e.g., Human Motor Area Template, or HMAT; Mayka et al., 2006), the latter being used by David Kemmerer in his in press paper in Brain and Language. David and I have already had some discussions about this topic, and David has included a section on this problem in his paper. My view is that functional neuroimaging studies require independent validation, with cytoarchitectonics being one method advocated by Eickhoff and colleagues. Mayka and colleagues seem to have the same view about their HMAT: “This template is not meant to serve as a gold standard for imaging studies of the motor system” (p. 1466). When using the HMAT, it should be recognised that the regions are delineated in Talairach atlas space, not MNI space. Consequently, coordinate conversion tools should be used to plot peaks in the latter space, see

In David’s paper, he also mentions a personal communication from Friedemann indicating that one of the peak maxima for the arm/hand in Hauk et al. (2004) was misreported, and that this might have influenced the conclusions in Postle et al. (2008) that were based upon a review of several studies. This raises the question of what we should consider as criteria for evidence of somatotopy: 1, 2, or 3 out of 3 effectors? My view is that all 3 are needed to make a convincing case.

The final issue pertains to the use of motor localisers, which Friedemann asserts makes the use of motor area maps irrelevant. I view this assertion as a non sequitur, as motor tasks activate large areas of cortex, not merely the motor regions. In Postle et al. (2008), we used the motor area maps to identify individuals’ activation peaks from a motor localiser task that involved viewing then executing a movement. In Hauk et al. (2004), participants read a verbal instruction that included the effector name while they performed the requisite movement in a block design experiment. Consequently, lexical processing was involved in both motor localiser and verb reading tasks, perhaps accounting for some overlap in the activation maps.


Eickhoff, S.B., et al. (2007). Assignment of functional activations to probabilistic cytoarchitectonic areas revisited. NeuroImage, 36, 511-521.

Mayka, M., et al. (2006). Locations and boundaries of the motor and premotor regions in functional brain imaging: a meta-analysis. NeuroImage 31, 1453–1474.