Predictive coding and how the dynamical Bayesian brain achieves specialization and integration

Authors note: this marks the first in a new series of journal-entry style posts in which I write freely about things I like to think about. The style is meant to be informal and off the cuff, building towards a sort of socratic dialogue. Please feel free to argue or debate any point you like. These are meant to serve as exercises in writing and thinking,  to improve the quality of both and lay groundwork for future papers. 

My wife Francesca and I are spending the winter holidays vacationing in the north Italian countryside with her family. Today in our free time our discussions turned to how predictive coding and generative models can accomplish the multimodal perception that characterizes the brain. To this end Francesca asked a question we found particularly thought provoking: if the brain at all levels is only communicating forward what is not predicted (prediction error), how can you explain the functional specialization that characterizes the different senses? For example, if each sensory hierarchy is only communicating prediction errors, what explains their unique specialization in terms of e.g. the frequency, intensity, or quality of sensory inputs? Put another way, how can the different sensations be represented, if the entire brain is only communicating in one format?

We found this quite interesting, as it seems straightforward and yet the answer lies at the very basis of predictive coding schemes. To arrive at an answer we first had to lay a little groundwork in terms of information theory and basic neurobiology. What follows is a grossly oversimplified account of the basic neurobiology of perception, which serves only as a kind of philosopher’s toy example to consider the question. Please feel free to correct any gross misunderstandings.

To begin, it is clear at least according to Shannon’s theory of information, that any sensory property can be encoded in a simple system of ones and zeros (or nerve impulses). Frequency, time, intensity, and so on can all be re-described in terms of a simplistic encoding scheme. If this were not the case then modern television wouldn’t work. Second, each sensory hierarchy presumably  begins with a sensory effector, which directly transduces physical fluctuations into a neuronal code. For example, in the auditory hierarchy the cochlea contains small hairs that vibrate only to a particular frequency of sound wave. This vibration, through a complex neuro-mechanic relay, results in a tonitopic depolarization of first order neurons in the spiral ganglion.

f1
The human cochlea, a fascinating neural-mechanic apparatus to directly transduce air vibrations into neural representations.

It is here at the first-order neuron where the hierarchy presumably begins, and also where functional specialization becomes possible. It seems to us that predictive coding should say that the first neuron is simply predicting a particular pattern of inputs, which correspond directly to an expected external physical property. To try and give a toy example, say we present the brain with a series of tones, which reliably increase in frequency at 1 Hz intervals. At the lowest level the neuron will fire at a constant rate if the frequency at interval n is 1 greater than the previous interval, and will fire more or less if the frequency is greater or less than this basic expectation, creating a positive or negative prediction error (remember that the neuron should only alter its firing pattern if something unexpected happens). Since frequency here is being signaled directly by the mechanical vibration of the cochlear hairs; the first order neuron is simply predicting which frequency will be signaled. More realistically, each sensory neuron is probably only predicting whether or not a particular frequency will be signaled – we know from neurobiology that low-level neurons are basically tuned to a particular sensory feature, whereas higher level neurons encode receptive fields across multiple neurons or features. All this is to say that the first-order neuron is specialized for frequency because all it can predict is frequency; the only afferent input is the direct result of sensory transduction. The point here is that specialization in each sensory system arises in virtue of the fact that the inputs correspond directly to a physical property.

f2
Presumably, first order neurons predict the presence or absence of a particular, specialized sensory feature owing to their input. Credit: wikipedia.

Now, as one ascends higher in the hierarchy, each subsequent level is predicting the activity of the previous. The first-order neuron predicts whether a given frequency is presented, the second perhaps predicts if a receptive field is activated across several similarly tuned neurons, the third predicts a particular temporal pattern across multiple receptive fields, and so on. Each subsequent level is predicting a “hyperprior” encoding a higher order feature of the previous level. Eventually we get to a level where the prediction is no longer bound to a single sensory domain, but instead has to do with complex, non-linear interactions between multiple features. A parietal neuron thus might predict that an object in the world is a bird if it sings at a particular frequency and has a particular bodily shape.

f3
The motif of hierarchical message passing which encompasses the nervous system, according the the Free Energy principle.

If this general scheme is correct, then according to hierarchical predictive coding functional specialization primarily arises in virtue of the fact that at the lowest level each hierarchy is receiving inputs that strictly correspond to a particular feature. The cochlea is picking up fluctuations in air vibration (sound), the retina is picking up fluctuations in light frequency (light), and the skin is picking up changes in thermal amplitude and tactile frequency (touch). The specialization of each system is due to the fact that each is attempting to predict higher and higher order properties of those low-level inputs, which are by definition particular to a given sensory domain. Any further specialization in the hierarchy must then arise from the fact that higher levels of the brain predict inputs from multiple sensory systems – we might find multimodal object-related areas simply because the best hyper-prior governing nonlinear relationships between frequency and shape is an amodal or cross-model object. The actual etiology of higher-level modules is a bit more complicate than this, and requires an appeal to evolution to explain in detail, but we felt this was a generally sufficient explanation of specialization.

Nonlinearity of the world and perception: prediction as integration

At this point, we felt like we had some insight into how predictive coding can explain functional specialization without needing to appeal to special classes of cortical neurons for each sensation. Beyond the sensory effectors, the function of each system can be realized simply by means of a canonical, hierarchical prediction of each layered input, right down to the point of neurons which predict which frequency will be signaled. However, something still was missing, prompting Francesca to ask – how can this scheme explain the coherent, multi-modal, integrated perception, which characterizes conscious experience?

Indeed, we certainly do not experience perception as a series of nested predictions. All of the aforementioned machinery functions seamlessly beyond the point of awareness. In phenomenology a way to describe such influences is as being prenoetic (before knowing; see also prereflective); i.e. things that influence conscious experience without themselves appearing in experience. How then can predictive coding explain the transition from segregated, feature specific predictions to the unified percept we experience?

f4
When we arrange sensory hierarchies laterally, we see the “markov blanket” structure of the brain emerge. Each level predicts the control parameters of subsequent levels. In this way integration arises naturally from the predictive brain.

As you might guess, we already hinted at part of the answer. Imagine if instead of picturing each sensory hierarchy as an isolated pyramid, we instead arrange them such that each level is parallel to its equivalent in the ‘neighboring’ hierarchy. On this view, we can see that relatively early in each hierarchy you arrive at multi-sensory neurons that are predicting conjoint expectations over multiple sensory inputs. Conveniently, this observation matches what we actually know about the brain; audition, touch, and vision all converge in tempo-parietal association areas.

Perceptual integration is thus achieved as easily as specialization; it arises from the fact that each level predicts a hyperprior on the previous level. As one moves upwards through the hierarchy, this means that each level predicts more integrated, abstract, amodal entities. Association areas don’t predict just that a certain sight or sound will appear, but instead encode a joint expectation across both (or all) modalities. Just like the fusiform face area predicts complex, nonlinear conjunctions of lower-level visual features, multimodal areas predict nonlinear interactions between the senses.

f5
A half-cat half post, or a cat behind a post? The deep convolutional nature of the brain helps us solve this and similar nonlinear problems.

It is this nonlinearity that makes predictive schemes so powerful and attractive. To understand why, consider the task the brain must solve to be useful. Sensory impressions are not generated by simple linear inputs; certainly for perception to be useful to an organism it must process the world at a level that is relevant for that organism. This is the world of objects, persons, and things, not disjointed, individual sensory properties. When I watch a cat walk behind a fence, I don’t perceive it as two halves of a cat and a fence post, but rather as a cat hidden behind a fence. These kinds of nonlinear interactions between objects and properties of the world are ubiquitous in perception; the brain must solve not for the immediately available sensory inputs but rather the complex hidden causes underlying them. This is achieved in a similar manner to a deep convolutional network; each level performs the same canonical prediction, yet together the hierarchy will extract the best-hidden features to explain the complex interactions that produce physical sensations. In this way the predictive brain summersaults the binding problem of perception; perception is integrated precisely because conjoint hypothesis are better, more useful explanations than discrete ones. As long as the network has sufficient hierarchical depth, it will always arrive at these complex representations. It’s worth noting we can observe the flip-side of this process in common visual illusions, where the higher-order percept or prior “fills in” our actual sensory experience (e.g. when we perceive a convex circle as being lit from above).

teaser-convexconcave-01
Our higher-level, integrative priors “fill in” our perception.

Beating the homunculus: the dynamic, enactive Bayesian brain

Feeling satisfied with this, Francesca and I concluded our fun holiday discussion by thinking about some common misunderstandings this scheme might lead one into. For example, the notion of hierarchical prediction explored above might lead one to expect that there has to be a “top” level, a kind of super-homunculus who sits in the prefrontal cortex, predicting the entire sensorium. This would be an impossible solution; how could any subsystem of the brain possibly predict the entire activity of the rest? And wouldn’t that level itself need to be predicted, to be realised in perception, leading to infinite regress? Luckily the intuition that these myriad hypotheses must “come together” fundamentally misunderstands the Bayesian brain.

Remember that each level is only predicting the activity of that before it. The integrative parietal neuron is not predicting the exact sensory input at the retina; rather it is only predicting what pattern of inputs it should receive if the sensory input is an apple, or a bat, or whatever. The entire scheme is linked up this way; the individual units are just stupid predictors of immediate input. It is only when you link them all up together in a deep network, that the brain can recapitulate the complex web of causal interactions that make up the world.

This point cannot be stressed enough: predictive coding is not a localizationist enterprise. Perception does not come about because a magical brain area inverts an entire world model. It comes about in virtue of the distributed, dynamic activity of the entire brain as it constantly attempts to minimize prediction error across all levels. Ultimately the “model” is not contained “anywhere” in the brain; the entire brain itself, and the full network of connection weights, is itself the model of the world. The power to predict complex nonlinear sensory causes arises because the best overall pattern of interactions will be that which most accurately (or usefully) explains sensory inputs and the complex web of interactions which causes them. You might rephrase the famous saying as “the brain is it’s own best model of the world”.

As a final consideration, it is worth noting some misconceptions may arise from the way we ourselves perform Bayesian statistics. As an experimenter, I formalize a discrete hypothesis (or set of hypotheses) about something and then invert that model to explain data in a single step. In the brain however the “inversion” is just the constant interplay of input and feedback across the nervous system at all levels. In fact, under this distributed view (at least according to the Free Energy Principle), neural computation is deeply embodied, as actions themselves complete the inferential flow to minimize error. Thus just like neural feedback, actions function as  ‘predictions’, generated by the inferential mechanism to render the world more sensible to our predictions. This ultimately minimises prediction error just as internal model updates do, albeit in a different ‘direction of fit’ (world to model, instead of model to world). In this way the ‘model’ is distributed across the brain and body; actions themselves are as much a part of the computation as the brain itself and constitute a form of “active inference”. In fact, if one extends their view to evolution, the morphological shape of the organism is itself a kind of prior, predicting the kinds of sensations, environments, and actions the agent is likely to inhabit. This intriguing idea will be the subject of a future blog post.

Conclusion

We feel this is an extremely exciting view of the brain. The idea that an organism can achieve complex intelligence simply by embedding a simple repetitive motif within a dynamical body seems to us to be a fundamentally novel approach to the mind. In future posts and papers, we hope to further explore the notions introduced here, considering questions about “where” these embodied priors come from and what they mean for the brain, as well as the role of precision in integration.

Questions? Comments? Feel like i’m an idiot? Sound off in the comments!

Further Reading:

Brown, H., Adams, R. A., Parees, I., Edwards, M., & Friston, K. (2013). Active inference, sensory attenuation and illusions. Cognitive Processing, 14(4), 411–427. http://doi.org/10.1007/s10339-013-0571-3
Feldman, H., & Friston, K. J. (2010). Attention, Uncertainty, and Free-Energy. Frontiers in Human Neuroscience, 4. http://doi.org/10.3389/fnhum.2010.00215
Friston, K., Adams, R. A., Perrinet, L., & Breakspear, M. (2012). Perceptions as Hypotheses: Saccades as Experiments. Frontiers in Psychology, 3. http://doi.org/10.3389/fpsyg.2012.00151
Friston, K., & Kiebel, S. (2009). Predictive coding under the free-energy principle. Philosophical Transactions of the Royal Society of London B: Biological Sciences, 364(1521), 1211–1221. http://doi.org/10.1098/rstb.2008.0300
Friston, K., Thornton, C., & Clark, A. (2012). Free-Energy Minimization and the Dark-Room Problem. Frontiers in Psychology, 3. http://doi.org/10.3389/fpsyg.2012.00130
Moran, R. J., Campo, P., Symmonds, M., Stephan, K. E., Dolan, R. J., & Friston, K. J. (2013). Free Energy, Precision and Learning: The Role of Cholinergic Neuromodulation. The Journal of Neuroscience, 33(19), 8227–8236. http://doi.org/10.1523/JNEUROSCI.4255-12.2013

 

Twitter Follow-up: Can MVPA Invalidate Simulation Theory?

Thanks to the wonders of social media, while I was out grocery shopping I received several interesting and useful responses to my previous post on the relationship between multivariate pattern analysis and simulation theory. Rather than try and fit my responses into 140 characters, I figured i’d take a bit more space here to hash them out. I think the idea is really enhanced by these responses, which point to several findings and features of which I was not aware. The short answer seems to be, no MVPA does not invalidate simulation theory (ST) and may even provide evidence for it in the realm of motor intentions, but that we might be able to point towards a better standard of evidence for more exploratory applications of ST (e.g. empathy-for-pain). An important point to come out of these responses as one might expect, is that the interpretation of these methodologies is not always straightforward.

I’ll start with Antonia Hamilton’s question, as it points to a bit of literature that speaks directly to the issue:

antonio_reply

Antonia is referring to this paper by Oosterhof and colleagues, where they directly compare passive viewing and active performance of the same paradigm using decoding techniques. I don’t read nearly as much social cognition literature as I used to, and wasn’t previously aware of this paper. It’s really a fascinating project and I suggest anyone interested in this issue read it at once (it’s open access, yay!). In the introduction the authors point out that spatial overlap alone cannot demonstrate equivalent mechanisms for viewing and performing the same action:

Numerous functional neuroimaging studies have identified brain regions that are active during both the observation and the execution of actions (e.g., Etzel et al. 2008; Iacoboni et al. 1999). Although these studies show spatial overlap of frontal and parietal activations elicited by action observation and execution, they do not demonstrate representational overlap between visual and motor action representations. That is, spatially overlapping activations could reflect different neural populations in the same broad brain regions (Gazzola and Keysers 2009; Morrison and Downing 2007; Peelen and Downing 2007b). Spatial overlap of activations per se cannot establish whether the patterns of neural response are similar for a given action (whether it is seen or performed) but different for different actions, an essential property of the “mirror system” hypothesis.”

They then go on to explain that while MVPA could conceivably demonstrate a simulation-like mechanism (i.e. a common neural representation for viewing/doing), several previous papers attempting to show just that failed to do so. The authors suggest that this may be due to a variety of methodological limitations, which they set out to correct for in their JNPhys publication. Oosterhof et al show that clusters of voxels located primarily in the intraparietal and superior temporal sulci encode cross-modal information, that is code similar information both when viewing and doing:

Click to go to PDF.
From Oosterhof et al, showing combined classification accuray for (train see, test do; train do, test see).

Essentially Oosterhof et al trained their classifier on one modality (see or do) , tested the classifier on the opposite modality in another session, and then repeated this procedure for all possible combinations of session and modality (while appropriately correcting for multiple comparisons). The map above represents the combined classification accuracy from both train-test combinations; interestingly in the supplementary info they show that the maps do slightly differ depend on what was trained:

Click to go to SI.
From supplementary info, A shows classifier trained on see, tested on do, B shows the opposite.

Oosterhof and colleagues also investigate the specificity of information for particular gestures in a second experiment, but for our purposes lets focus on just the first. My first thought is that this does actually provide some evidence for a simulation theory of understanding motor intentions. Clearly there is enough information in each modality to accurately decode the opposite modality: there are populations of neurons encoding similar information both for action execution and perception. Realistically I think this has to be the minimal burden of proof needed to consider an imaging finding to be evidence for simulation theory. So the results of Oosterhof et al do provide supporting evidence for simulation theory in the domain of motor intentions.

Nonetheless, the results also strengthen the argument that more exploratory extentions of ST (like empathy-for-pain) must be held to a similar burden of proof before generalization in these domains is supported. Simply showing spatial overlap is not evidence of simulation, as Oosterhof themselves argue. I think it is interesting to note the slight spatial divergence between the two train-test maps (see on do, do on see). While we can obviously identify voxels encoding cross-modality information, it is interesting that those voxels do not subsume the entirety of whatever neural computation relates these two modalities; each has something unique to predict in the other. I don’t think that observation invalidates simulation theory, but it might suggest an interesting mechanism not specified in the ‘vanilla’ flavor of ST. To be extra boring, it would be really nice to see an independent replication of this finding, since as Oosterhof themselves point out, the evidence for cross-modal information is inconsistent across studies. Even though the classifier performs well above chance in this study,  it is also worth noting that the majority of surviving voxels in their study show somewhere around 40-50% classification accuracy, not exactly gangbusters. It would be interesting to see if they could identify voxels within these regions that selectively encode only viewing or performing; this might be evidence for a hybrid-theory account of motor intentions.

leoreply

Leonhard’s question is an interesting one that I don’t have a ready response for. As I understand it, the idea is that demonstrating no difference of patterns between a self and other-related condition (e.g. performing an action vs watching someone else do it) might actually be an argument for simulation, since this could be caused by that region using isomorphic computations for both conditions. This an interesting point – i’m not sure what the status of null findings is in the decoding literature, but this merits further thought.

The next two came from James Kilner and Tal Yarkoni. I’ve put them together as I think they fall under a more methodological class of questions/comments and I don’t feel quite experienced enough to answer them- but i’d love to hear from someone with more experience in multivariate/multivoxel techniques:

kilner_reply

talreply

James Kilner asks about the performance of MVPA in the case that the pattern might be spatially overlapping but not identical for two conditions. This is an interesting question and i’m not sure I know the correct answer; my intuition is that you could accurately discriminate both conditions using the same voxels and that this would be strong evidence against a simple simulation theory account (spatial overlap but representational heterogeneity).

Here is more precise answer to James’ question from Sam Schwarzkopf, posted in the comments of the original post:

2. The multivariate aspect obviously adds sensitivity by looking at pattern information, or generally any information of more than one variable (e.g. voxels in a region). As such it is more sensitive to the information content in a region than just looking at the average response from that region. Such an approach can reveal that region A contains some diagnostic information about an experimental variable while region B does not, even though they both show the same mean activation. This is certainly useful knowledge that can help us advance our understanding of the brain – but in the end it is still only one small piece in the puzzle. And as both Tal and James pointed out (in their own ways) and as you discussed as well, you can’t really tell what the diagnostic information actually represents.
Conversely, you can’t be sure that just because MVPA does not pick up diagnostic information from a region that it therefore doesn’t contain any information about the variable of interest. MVPA can only work as long as there is a pattern of information within the features you used.

This last point is most relevant to James’ comment. Say you are using voxels as features to decode some experimental variable. If all the neurons with different tuning characteristics in an area are completely intermingled (like orientation-preference in mouse visual cortex for instance) you should not really see any decoding – even if the neurons in that area are demonstrably selective to the experimental variable.

In general it is clear that the interpretation of decoded patterns is not straightforward- it isn’t clear precisely what information they reflect, and it seems like if a region contained a totally heterogeneous population of neurons you wouldn’t pick up any decoding at all. With respect to ST,  I don’t know if this completely invalidates our ability to test predictions- I don’t think one would expect such radical heterogeneity in a region like STS, but rather a few sub-populations responding selectively to self and other, which MVPA might be able to reveal. It’s an important point to consider though.

Tal’s point is an important one regarding the different sources of information that GLM and MVPA techniques pick up. The paper he refers to by Jimura and Poldrack set out to investigate exactly this by comparing the spatial conjunction and divergent sensitivity of each method. Importantly they subtracted the mean of each beta-coefficient from the multivariate analysis to insure that the analysis contained only information not in the GLM:

pold_mvpa

As you can see in the above, Jimura and Poldrack show that MVPA picks up a large number of voxels not found in the GLM analysis. Their interpretation is that the GLM is designed to pick up regions responding globally or in most cases to stimulation, whereas MVPA likely picks up globally distributed responses that show variance in their response. This is a bit like the difference between functional integration and localization; both are complementary to the understanding of some cognitive function. I take Tal’s point to be that the MVPA and GLM are sensitive to different sources of information and that this blurs the ability of the technique to evaluate simulation theory- you might observe differences between the two that would resemble evidence against ST (different information in different areas) when in reality you would be modelling altogether different aspects of the cognition. edit: after more discussion with Tal on Twitter, it’s clear that he meant to point out the ambiguity inherent in interpreting the predictive power of MVPA; by nature these analyses will pick up a lot of confounding a causal noise- arousal, reaction time, respiration, etc, which would be excluded in a GLM analysis. So these are not necessarily or even likely to be “direct read-outs” of representations, particularly to the extent that such confounds correlate with the task. See this helpful post by neuroskeptic for an overview of one recent paper examining this issue. See here for a study investigating the complex neurovascular origins of MVPA for fMRI. 

Thanks sincerely for these responses, as it’s been really interesting and instructive for me to go through these papers and think about their implications. I’m still new to these techniques and it is exciting to gain a deeper appreciation of the subtleties involved in their interpretation. On that note, I must direct you to check out Sam Schwarzkopf’s excellent reply to my original post. Sam points out some common misunderstandings (of which I am perhaps guilty of several) regarding the interpretation of MVPA/decoding versus GLM techniques, arguing essentially that they pick up much of the same information and can both be considered ‘decoding’ in some sense, further muddying their ability to resolves debates like that surrounding simulation theory.

Will multivariate decoding spell the end of simulation theory?

Decoding techniques such as multivariate pattern analysis (MVPA) are hot stuff in cognitive neuroscience, largely because they offer a tentative promise of actually reading out the underlying computations in a region rather than merely describing data features (e.g. mean activation profiles). While I am quite new to MVPA and similar machine learning techniques (so please excuse any errors in what follows), the basic process has been explained to me as a reversal of the X and Y variables in a typical general linear model. Instead of specifying a design matrix of explanatory (X) variables and testing how well those predict a single independent (Y) variable (e.g. the BOLD timeseries in each voxel), you try to estimate an explanatory variable (essentially decoding the ‘design matrix’ that produced the observed data) from many Y variables, for example one Y variable per voxel (hence the multivariate part). The decoded explanatory variable then describes (BOLD) responses in way that can vary in space, rather than reflecting an overall data feature across a set of voxels such as mean or slope. Typically decoding analyses proceed in two steps, one in which you train the classifier on some set of voxels and another where you see how well that trained model can classify patterns of activity in another scan or task. It is precisely this ability to detect patterns in subtle spatial variations that makes MVPA an attractive technique- the GLM simply doesn’t account for such variation.

The implicit assumption here is that by modeling subtle spatial variations across a set of voxels, you can actually pick up the neural correlates of the underlying computation or representation (Weil and Rees, 2010, Poldrack, 2011). To illustrate the difference between an MVPA and GLM analysis, imagine a classical fMRI experiment where we have some set of voxels defining a region with a significant mean response to your experimental manipulation. All the GLM can tell us is that in each voxel the mean response is significantly different from zero. Each voxel within the significant region is likely to vary slightly in its actual response- you might imagine all sorts of subtle intensity variations within a significant region- but the GLM essentially ignores this variation. The exciting assumption driving interest in decoding is that this variability might actually reflect the activity of sub-populations of neurons and by extension, actual neural representations. MVPA and similar techniques are designed to pick out when these reflect a coherent pattern; once identified this pattern can be used to “predict” when the subject was seeing one or another particular stimulus. While it isn’t entirely straightforward to interpret the patterns MVPA picks out as actual ‘neural representations’, there is some evidence that the decoded models reflect a finer granularity of neural sub-populations than represented in overall mean activation profiles (Todd, 2013; Thompson 2011).

Professor Xavier applies his innate talent for MVPA.
Professor Xavier applies his innate talent for MVPA.

As you might imagine this is terribly exciting, as it presents the possibility to actually ‘read-out’ the online function of some brain area rather than merely describing its overall activity. Since the inception of brain scanning this has been exactly the (largely failed) promise of imaging- reverse inference from neural data to actual cognitive/perceptual contents. It is understandable then that decoding papers are the ones most likely to appear in high impact journals- just recently we’ve seen MVPA applied to dream states, reconstruction of visual experience, and pain experience all in top journals (Kay et al., 2008, Horikawa et al., 2013, Wager et al., 2013). I’d like to focus on that last one for the remainer of this post, as I think we might draw some wide-reaching conclusions for theoretical neuroscience as a whole from Wager et al’s findings.

Francesca and I were discussing the paper this morning- she’s working on a commentary for a theoretical paper concerning the role of the “pain matrix” in empathy-for-pain research. For those of you not familiar with this area, the idea is a basic simulation-theory argument-from-isomorphism. Simulation theory (ST) is just the (in)famous idea that we use our own motor system (e.g. mirror neurons) to understand the gestures of others. In a now infamous experiment Rizzolatti et al showed that motor neurons in the macaque monkey responded equally to their own gestures or the gestures of an observed other (Rizzolatti and Craighero, 2004). They argued that this structural isomorphism might represent a general neural mechanism such that social-cognitive functions can be accomplished by simply applying our own neural apparatus to work out what was going on for the external entity. With respect to phenomena such empathy for pain and ‘social pain’ (e.g. viewing a picture of someone you broke up with recently), this idea has been extended to suggest that, since a region of networks known as “the pain matrix” activates similarly when we are in pain or experience ‘social pain’, that we “really feel” pain during these states (Kross et al., 2011) [1].

In her upcoming commentary, Francesca points out an interesting finding in the paper by Wager and colleagues that I had overlooked. Wager et al apply a decoding technique in subjects undergoing painful and non-painful stimulation. Quite impressively they are then able to show that the decoded model predicts pain intensity in different scanners and various experimental manipulations. However they note that the model does not accurately predict subject’s ‘social pain’ intensity, even though the subjects did activate a similar network of regions in both the physical and social pain tasks (see image below). One conclusion from these findings it that it is surely premature to conclude that because a group of subjects may activate the same regions during two related tasks, those isomorphic activations actually represent identical neural computations [2]. In other words, arguments from structural isomorpism like ST don’t provide any actual evidence for the mechanisms they presuppose.

Figure from Wager et al demonstrating specificity of classifier for pain vs warmth and pain vs rejection. Note poor receiver operating curve (ROC) for 'social pain' (rejecter vs friend), although that contrast picks out similar regions of the 'pain matrix'.
Figure from Wager et al demonstrating specificity of classifier for pain vs warmth and pain vs rejection. Note poor receiver operating curve (ROC) for ‘social pain’ (rejecter vs friend), although that contrast picks out similar regions of the ‘pain matrix’.

To me this is exactly the right conclusion to take from Wager et al and similar decoding papers. To the extent that the assumption that MVPA identifies patterns corresponding to actual neural representations holds, we are rapidly coming to realize that a mere mean activation profile tells us relatively little about the underlying neural computations [3]. It certainly does not tell us enough to conclude much of anything on the basis that a group of subjects activate “the same brain region” for two different tasks. It is possible and even likely that just because I activate my motor cortex when viewing you move, I’m doing something quite different with those neurons than when I actually move about. And perhaps this was always the problem with simulation theory- it tries to make the leap from description (“similar brain regions activate for X and Y”) to mechanism, without actually describing a mechanism at all. I guess you could argue that this is really just a much fancier argument against reverse inference and that we don’t need MVPA to do away with simulation theory. I’m not so sure however- ST remains a strong force in a variety of domains. If decoding can actually do away with ST and arguments from isomorphism or better still, provide a reasonable mechanism for simulation, it’ll be a great day in neuroscience. One thing is clear- model based approaches will continue to improve cognitive neuroscience as we go beyond describing what brain regions activate during a task to actually explaining how those regions work together to produce behavior.

I’ve curated some enlightening responses to this post in a follow-up – worth checking for important clarifications and extensions! See also the comments on this post for a detailed explanation of MVPA techniques. 

References

Horikawa T, Tamaki M, Miyawaki Y, Kamitani Y (2013) Neural Decoding of Visual Imagery During Sleep. Science.

Kay KN, Naselaris T, Prenger RJ, Gallant JL (2008) Identifying natural images from human brain activity. Nature 452:352-355.

Kross E, Berman MG, Mischel W, Smith EE, Wager TD (2011) Social rejection shares somatosensory representations with physical pain. Proceedings of the National Academy of Sciences 108:6270-6275.

Poldrack RA (2011) Inferring mental states from neuroimaging data: from reverse inference to large-scale decoding. Neuron 72:692-697.

Rizzolatti G, Craighero L (2004) The mirror-neuron system. Annu Rev Neurosci 27:169-192.

Thompson R, Correia M, Cusack R (2011) Vascular contributions to pattern analysis: Comparing gradient and spin echo fMRI at 3T. Neuroimage 56:643-650.

Todd MT, Nystrom LE, Cohen JD (2013) Confounds in Multivariate Pattern Analysis: Theory and Rule Representation Case Study. NeuroImage.

Wager TD, Atlas LY, Lindquist MA, Roy M, Woo C-W, Kross E (2013) An fMRI-Based Neurologic Signature of Physical Pain. New England Journal of Medicine 368:1388-1397.

Weil RS, Rees G (2010) Decoding the neural correlates of consciousness. Current opinion in neurology 23:649-655.


[1] Interestingly this paper comes from the same group (Wager et al) showing that pain matrix activations do NOT predict ‘social’ pain. It will be interesting to see how they integrate this difference.

[2] Nevermind the fact that the ’pain matrix’ is not specific for pain.

[3] With all appropriate caveats regarding the ability of decoding techniques to resolve actual representations rather than confounding individual differences (Todd et al., 2013) or complex neurovascular couplings (Thompson et al., 2011).