Dear Greg and David:
Thank you for your comments on our 2009 paper (Fedorenko & Kanwisher, 2009, Language and Linguistics Compass). As you stated in your post on August 24, 2010, and as became even clearer from the email exchange that took place the following day, you agree with Nancy Kanwisher and me on the importance of taking into consideration anatomical and functional variability in fMRI research in any domain, including language. This was the main point of our article, so on the important points we agree. We will here focus on the points where there is some disagreement between us.
1. How much do we, as a field, know about functional specialization of language-sensitive brain regions?
Although different researchers in the fields of psycho- and neuro-linguistics may vary in how satisfied they are with the progress that has been made in understanding the brain basis of language, we don't think anyone would argue that we have arrived at a complete understanding of what computations each language-sensitive brain region performs and/or what representations it stores/manipulates. So, given that we can probably agree on this, the question becomes: how much do we actually know, as a field, about the brain basis of language in general, and about the questions of functional specialization of language-sensitive brain regions, in particular.
As you point out, we have argued that two fundamental questions concerning the functional architecture of the language system remain unanswered:
• Do language-sensitive brain regions each support one particular aspect of language, or do they support multiple linguistic processes?
• Do language-sensitive regions specialize for (some aspect of) language processing, or do they also support some non-linguistic processes (e.g., general working memory, arithmetic processing, music, action representation, etc.)?
You implied that our pessimistic assessment is due to ignorance. In fact, we are aware of the relevant literature, but we think that the findings that you refer to, among others, either do not pertain to or conclusively answer the questions of functional specificity.
Let us first clarify what we mean by “functional specificity”, because perhaps some of the confusion stems from different interpretations of the issues at hand. Here is our working definition: A brain region R is specialized for cognitive function x if this region (i) is engaged in tasks that rely on cognitive function x, and (ii) is not engaged in tasks that do not rely on cognitive function x.
It is important to clarify the difference between a brain region supporting a particular cognitive function and a brain region selectively supporting that cognitive function. A demonstration of a relationship between brain region R and cognitive function x tells us little about R’s specificity for x. For example, various working memory tasks consistently activate dorsolateral prefrontal cortical regions, across studies and labs. However, this does not mean that these regions are selectively engaged in working memory; instead, these regions appear to support a wide range of demanding cognitive tasks (e.g., Duncan, 2001, Nat Rev Neurosci Duncan, 2010, TiCS). We elaborate on this issue – with respect to the claim that a portion of Broca’s area is selectively engaged in one aspect of syntax (e.g., Grodzinsky & Santi, 2008, TiCS) – in our response to Grodzinsky (Fedorenko & Kanwisher, submitted; see also Willems & Hagoort, 2009, TiCS, for similar arguments).
Now, let’s turn to your claims.
Concerning within-language specificity, you said:
“… there is strong evidence from functional imaging regarding the involvement of distinct cortical regions/circuits in phonemic (STS), lexical-semantic (MTG), prosodic (anterior dorsal STG), and higher-level combinatorial processes (anterior temporal/inferior frontal regions). Additional circuits have been delineated that support auditory-motor integration and auditory/phonological short-term memory. Here are some relevant reviews of this literature: Binder et al., 2000; Hickok and Poeppel, 2007; Indefrey and Levelt, 2004.”
Concerning domain specificity of language-sensitive regions, you said:
“Regarding question two, several studies have clearly identified voice-specific responses in the STG (Belin et al. 2000), higher-level speech specific responses in the STS (Scott et al. 2000; Okada et al. 2010), and even what we might as well call the anterior temporal lobe sentence area, given how selective it is to the perception of sentence-level stimulation (Humphries et al., 2006; Humphries et al., 2005; Humphries et al., 2001; Rogalsky and Hickok, 2009; Vandenberghe et al., 2002). But more importantly, some of us have moved beyond the specificity issue with the aim of trying to identify the circuits and computations involved in a given process whether or not it is special to speech.”
Then later you pointed out that we have not cited several important findings from the recent literature. Of relevance to the questions of functional specificity are the following (NB: (i) and (iii) are not directly relevant to this question, so we will not discuss those here):
“(ii) convergence on the involvement of the STS in phonemic level processes in speech perception (Leibenthal et al., 2005; Scott & Johnsrude, 2003; Hickok & Poeppel, 2007)”;
“(iv) recent progress in mapping the circuit that supports sensory-motor integration in speech processing (Golfinopoulus et al. 2009; Hickok et al., 2009)”;
“(v) progress in understanding the basis of hemispheric asymmetries for acoustic and phonemic processing in auditory cortex (Boemio, et al. 2005); Zatorre, et al. 2002)
“(vi) convergence on the idea that anterior temporal regions support some aspect of sentence-level processing (the linguistic equivalent of the FFA)”;
“(vii) convergence on the relation between sensory-motor circuits and phonological short-term memory (Buchsbaum et al. 2008; Postle, 2006)”.
Although we are familiar with the findings and papers you mention, we don’t think that – with respect to any of the brain regions above – we can conclusively say: this region exclusively performs cognitive function x, or cognitive functions x1-xn (if a region is truly multifunctional). This is not to say that some of the studies you mention, as well as dozens of other studies, have not narrowed down the space of possible hypotheses for the function(s) of various regions.
Let’s consider one of your examples above: the claim that “anterior temporal regions support some aspect of sentence-level processing”; this area, according to you, is “selective […] to the perception of sentence-level stimulation”, and you even go as far as calling it “the linguistic equivalent of the FFA”. The papers you cite in support of this claim indeed quite strongly suggest that this region is engaged in processing sentence-level material. However, in order to argue that this region selectively supports sentence-level understanding, more work is needed. In particular, language has been argued to share various properties with a number of non-language tasks (e.g., arithmetic, musical processing, action representation, general executive functions, aspects of social cognition, etc.). Furthermore, tasks in some of these domains have been shown to activate regions in/around anterior temporal cortex (just to provide a few examples: amodal conceptual processing – e.g., Patterson et al., 2007, Nat Rev Neurosci; memory for music – e.g., Peretz, 1996, J of Cog Neurosci; aspects of social cognition – e.g., Zahn et al., 2007, PNAS; Ross & Olson, 2010, Neuroimage; face processing – e.g., Sugiura et al., 2001, Neuroimage). Perhaps these different non-language tasks activate different portions of the ATL than sentence-level understanding, and perhaps you can dismiss some of these claims / findings on the basis of some design, methodological or analysis flaws. Nevertheless, in order to convince the world that anterior temporal regions are specialized for some aspect of sentence-level processing (as you seem to argue), you would have to show that those regions do not respond to any of the tasks that have been shown to activate regions in/around ATL (as e.g., Kanwisher and colleagues have done for the FFA and other ventral visual stream regions).
Another finding you mention is “the involvement of the STS in phonemic level processes in speech perception”. As with the role of the ATL structures in sentence-level understanding, we agree with you that there is strong evidence for the engagement of some portions of STS in phonemic level processes. Perhaps this is even the strongest case of functional specialization in language-related brain regions since across several studies (including the ones you cite) a stronger response to speech sounds has been demonstrated relative to several non-speech auditory control conditions (cf. Leech et al., 2009, J Neurosci, for a claim that speech-sound-sensitive areas in STS are not domain specific but rather subserve the processing of acoustically complex sounds more generally and only show higher responses to speech sounds due to more extensive experience with those sounds). However, as with the ATL structures, we think more work is needed. Within language, some studies have argued for the sensitivity of STS regions (including the posterior portion of it) to high-level linguistic manipulations (see e.g., Caplan, 2007, Language & Linguistics Compass, for a review of some such studies; see also Friederici et al., 2010, Hum Brain Mapping, among others). Furthermore, a number of non-linguistic tasks have been shown to activate regions in/around STS, such as action representation (e.g., Vander Wyk et al., 2009, Psych Sci) or aspects of social cognition (e.g., Allison et al., 2000, TiCS), among others (see Hein & Knight, 2008, J Cog Neurosci, for a recent review). Again, as with the ATL structures, these different tasks may activate regions that are non-overlapping with the speech-sound-sensitive regions. However, this needs to be demonstrated empirically.
In summary, in order to make claims about functional specificity of some region R for a cognitive function x, it is necessary to examine R’s response to a wide range of linguistic and non-linguistic tasks that have either been (a) argued to share some properties with x on theoretical grounds, and/or (b) shown to activate cortex in/around R. (Group-based methods can be used for addressing these questions, but individual-subject analyses are particularly well suited as they will be most likely to uncover dissociations if such are present (Nieto-Castanon, Fedorenko & Kanwisher, in prep.))
Nevertheless, despite the fact that questions of functional specificity have not yet been “nailed” for any of the language-sensitive regions, you are right that previous work has contributed significantly to narrowing down the space of possible hypotheses for the functions of many of these regions (and perhaps this is what you wanted us to acknowledge more in our 2009 paper). Our 2009 paper was not intended to be a complete review of the relevant literature. Instead, our approach in that paper (as well as in the 2010 J Neurophys paper) has been to fairly broadly state the research questions that, in our opinion, remain open as of yet (i.e., the questions of functional specificity of language-sensitive regions). This approach has its limitations: some researchers may feel neglected and/or not properly acknowledged for their contributions to the field. We apologize for any such omissions.
2. Understanding the degree of functional specialization of a region is critical to discerning the computations it supports and/or the representations it stores/manipulates.
“But more importantly, some of us have moved beyond the specificity issue with the aim of trying to identify the circuits and computations involved in a given process whether or not it is special to speech.”
We don't understand this assertion. We also would like to understand the computations that different language-sensitive brain regions perform. However, understanding whether / how specialized a particular region is for language (or a particular aspect of linguistic processing), is essential for understanding the region's computations, so the juxtaposition is confusing.
There are at least three possible research strategies for discerning the nature of the computations/representations in language-sensitive brain regions. One could start by characterizing the language system (i.e., all the language-sensitive brain regions) in terms of their within-language specificity: figuring out whether different regions specialize for particular aspects of language vs. support multiple aspects of language, and understanding how these regions work together to enable language production /comprehension. Then one could use this information to further examine whether these regions support any non-linguistic processes. Alternatively, one could start by characterizing language-sensitive brain regions in terms of their domain specificity: figuring out which, if any, non-linguistic cognitive processes different language-sensitive regions support. Then one could use this information to further examine the precise role of these regions in linguistic processing. Finally, one could simultaneously tackle both questions – (i) within-language specificity, and (ii) domain specificity – and treat the findings with respect to these questions as mutually informative. We are adopting the third approach and hope to use knowledge accumulated in both of these lines of research to draw inferences about the function(s) of each language-sensitive region.
(Of course, fMRI is also not the only method that is needed to answer these questions definitively: high-temporal-resolution methods will be critical for potentially dissociating computations that may be performed by the same cortical tissue but occur at different time-scales.)
Somewhat relatedly, in another one of your posts (http://www.talkingbrains.org/2008/09/brodmann-areas-and-localization-in.html), you said the following:
“Ted Jones provides an instructive reminder in this new age of localization-based neuroscience: “No cortical area is an isolated entity in which a single function is represented. Nor, contrary to many current views, does it merely form one step in a hierarchy of areas proceeding onwards and upwards to some defined or imagined higher function. While there are definite streams of cortico–cortical connections that proceed in identifiable ways from area to area in the cortex, no area is without feedback connections and no area is without re-entrant connections from the thalamus.””
Although it is certainly true that any given brain region is connected to numerous other brain regions, we do not agree with the assertion that “no cortical area is an […] entity in which a single function is represented”. Several regions in the ventral visual stream have been shown to be exquisitely specialized for processing visual stimuli of a particular class (see e.g., Kanwisher, 2010, PNAS, for a recent overview). Furthermore, Saxe and colleagues have shown that a region in the right temporo-parietal junction selectively responds to stimuli that require us to think about what another person is thinking (e.g., Saxe & Powell, 2006, Psych Sci, and many other papers; see the publications section on the SaxeLab’s website: http://saxelab.mit.edu/publications.php). It is possible that none of the language-sensitive brain regions will show this degree of functional specialization. However, this question deserves to be investigated in the most rigorous way possible given its centrality to understanding human cognition in general.
Given numerous reports of what look like highly specialized linguistic deficits following focal brain damage, we tend to think that at least some degree of functional specialization is bound to be present in the language system. However, we are not on a quest to demonstrate functional specialization of some brain region for language (or for a particular aspect of language). Instead, we want to understand the functional response profiles of different language-sensitive brain regions, and the methods we advocate are perfectly suited for this task. Only by understanding how a region responds to a wide range of cognitive tasks can we begin to make inferences about the computations it performs and the representations that it stores/manipulates.
3. A couple of comments on “methodological schooling”.
“This is frankly a pathetic summary of the state of the field and a pretentious starting point for the methodological schooling that F&K provide in the following sections of their paper.”
“Methodological schooling” is not our intention. A few points are worth making here.
a) Many brain imaging methods have a place in investigations of the brain basis of language.
We tried to make this clear in our 2010 paper and in our response to Grodzinsky (Fedorenko & Kanwisher, submitted) and will keep trying to get this point across in our future talks and papers: we strongly believe that individual-subject analyses (both the “basic” subject-specific fROI analyses as well as some more sophisticated individual-subject analyses we’ve been developing) can go a long way in providing clearer answers to some outstanding questions. However, we are not arguing for abandoning the traditional, group-based, methods (or for ignoring the knowledge we have gained from those methods). Rather we think that a multi-pronged approach to the study of language – using all the tools available to us – is what’s needed for faster progress in the field.
b) How novel are functional localizers in language research?
We do realize and explicitly acknowledge (e.g., Fedorenko et al., 2010, J Neurophys, p. 1178) that some researchers have used individual subject analyses previously (we apologize for not including your studies in this list; we will do that in the future). However, as we note in the paper, to the best of our knowledge, prior to our 2010 paper (and efforts in the clinical literature to localize language-sensitive cortex using fMRI, which have not been adopted in the non-clinical literature), no contrast aimed at localizing language-sensitive brain regions has been validated, i.e., shown to pick out the same regions reliably within and across subjects. Without independent validation of the functional localizer contrasts, it may be difficult to interpret the responses of the ROIs to the critical condition of interest. What are needed are not “ad hoc” localizers that are different across different studies / labs but standardized ones that have been shown to effectively identify the target regions. For example, the localizer that you used in Rogalsky & Hickok (2009, Cereb Cortex) – a contrast between sentences and word lists – identifies only a region in the ATL and does not find any posterior temporal or frontal regions, so does not manage to capture many regions known to be engaged in sentence processing. A key goal of our 2010 J Neurophys paper was to test and establish the robustness of a specific localizer that successfully identifies most of these regions and hence will be of widespread use for many future studies. (Note that although we use the contrast between sentences and nonwords as our main contrast, the contrast between sentences and word lists in our localizer identifies very similar regions to those identified with the sentences > nonword lists contrast, including left frontal and temporal / temporo-parietal regions.)
Another contribution of our work is in developing a method for objectively and efficiently delineating the borders of functional ROIs in individual subjects, the group-constrained subject-specific (GSS) method (for more info see: http://web.mit.edu/evelina9/www/funcloc.html; see Julian et al., in prep., for validation of this method on well characterized high-level visual regions).
c) We invite others to try our methods.
As we mentioned in an email to you, by developing a way to quickly and reliably identify language-sensitive regions, we are not trying to say, “You have all done this wrong; see how we can do it better", but rather, we want to relay something along the lines of, “There may be some limitations in the methods that are currently in use in the vast majority of fMRI studies of language; here is a new approach: let's see how we, as a field, can put our heads together and harness the power of individual subject analyses – by developing a series of standard language localizers and using them consistently across studies and labs – to bring more clarity to where it's needed”.
By making all our tools publicly available (http://web.mit.edu/evelina9/www/funcloc.html), we hope that other researchers will join us in adopting the subject-specific approach and that, as a field, we can develop a set of standard localizers for various aspects of language that can be used consistently across studies and labs. We can then – in a joint effort – systematically characterize a set of key brain regions that enable proper linguistic functioning by examining their response to a wide range of linguistic and non-linguistic stimuli and tasks, thus deriving detailed functional profiles for each region.
We would like to conclude by commenting on the following assertion you make at the end of your post:
“individual subject analyses would be helpful, but it is not a magic bullet (e.g., task selection is more important in my view)”
First, we agree that task selection is highly important in characterizing functional profiles of language-sensitive (or other) regions. However, we don’t have to choose between good task design and the use of ROIs: we can do both.
Second, we are advocating individual subject analyses, and we strongly believe that such analyses will bring a clearer understanding of the brain basis of language. We have not shown this yet, so there is no point to argue against this. So far, we developed a method and several new analysis tools that now enable us to pursue questions of functional specificity with rigor that has not so far been possible, because (i) group-based methods are guaranteed to underestimate specificity, and (ii) without a standard localizer used to define ROIs across studies / labs, it is difficult to compare activations across studies / labs, because one can never – with certainty – determine whether an activation peak in one study reflects the activity of the same region as the activation peak in another study, or of a different / nearby region. Of course, it remains to be seen what this method can bring to the field in terms of our understanding of the language architecture. We think it's too early to make pronouncements like the one you make above. Why not give this method a chance? We invite you and others in the field to join our efforts.
Ev Fedorenko and Nancy Kanwisher
P.S. Our response to Grodzinsky’s critique of our 2009 paper will soon be available from Ev’s website (http://web.mit.edu/evelina9/www/).
P.P.S. One last comment – not related to the issues above – remains to be made. You made a parallel between the current debate and another debate that I (Ev) was recently involved in: the debate about whether quantitative methods are needed in linguistics research (Gibson & Fedorenko, 2010, TiCS; Gibson & Fedorenko, in press, LCP). These two debates are quite different, in spite of some superficial similarities. Unlike brain imaging research on language, where a choice of an analysis method may be driven by the nature of the research question and sometimes by researchers’ preferences, the use of quantitative methods in linguistics it is not a matter of a research question or preference/opinion. Quantitative methods should be used in linguistics, as they are used in every other branch of cognitive science, and science more generally.