Friday, March 26, 2010

Self-destruction of the mirror neuron theory of action understanding

Rizzolatti & Sinigaglia's new Nature Reviews Neuroscience paper on the mirror system is effectively an admission that the mirror neuron theory of action understanding is wrong. The original is idea was interesting: we understand actions by mirroring those actions in our own motor system. But this is no longer the case according to R&S:

By matching individual movements, mirror processing provides a representation of body part movement that might serve various functions (for example, imitation), but is devoid of any specific cognitive importance per se. p. 269

Instead, understanding comes from matching higher-order representations of the goals of the action (the quote above continues immediately):

By contrast, through matching the goal of the observed motor act with a motor act that has the same goal, the observer is able to understand what the agent is doing. p. 269

This goal-matching, according to R&S is quite independent of any specific motor act.
...among the neurons in various areas that become active during action observation, only those that can encode the goal of the motor behaviour of another individual with the greatest degree of generality can be considered to be crucial for action understanding... Indeed, parieto-frontal mirror neurons encode the goal of observed motor acts regardless of whether they are performed with the mouth, the hand or even with tools. p. 269

So, mirror neurons, those cells that fire during specific actions such as grasping-with-the-hand and while watching the same specific action -- the very cells that got everyone SO excited -- are not involved in action understanding. Rather, according to R&S, action understanding is achieved by cells that do not code for actions at all, but something higher level, goals/intentions.

It's worth noting that R&S directly contradict themselves in the sidebar definition of "Mirror-based action understanding":

The comprehension of an observed action based on the activation of a motor programme in the observer’s brain. p. 265

A motor program presumably controls a specific action, such as grasping-with-the-hand, not an action-independent goal or intention.

Is the mirror system needed for coding all goals and intentions? No, according to R&S:
This does not mean that the parieto-frontal mirror mechanism mediates all varieties of intention understanding. p. 271

But they want to say that the system IS needed for coding motor goals/intentions. They illustrate with an example:

Mary is interacting with an object (for example, a cup). According to how she is grasping the cup, we can understand why she is doing it (for example, to drink from it or to move it). This kind of understanding can be mediated by the parieto-frontal mirror mechanism by virtue of its motor chain organization. p. 271

It is true that we can make limited inferences about what Mary's intentions are by the way she grasps the cup. But (i) these inferences are underdetermined by the movement and (ii) motor experience with grasping cups is not necessary for making these inferences. Regarding (i), if Mary grasps the handle, rather than pushing the side with her fingertips, we may infer that she intends to drink rather than move. However, Mary could just as well be moving the cup, picking it up to put it in the sink, or picking it up to give to someone. Regarding (ii), simply having perceptual experience with grasping for drinking versus pushing for moving actions will result in the same inferential ability even without motor experience with cups.

Acknowledging the point that action understanding does not require the motor system, R&S site several studies that show that understanding can be achieved without the mirror system. They conclude,

These data indicate that the recognition of the motor behaviour of others can rely on the mere processing of its visual aspects. p. 270

So, to summarize:

1. Cells that mirror specific actions (i.e., congruent mirror neurons), don't support action understanding.
2. The real work of action understanding is done by cells that abstract away from actions and instead code goals and intentions.
3. Intentions can be coded outside the mirror/motor system.
4. The recognition of actions of others can be achieved outside the mirror/motor system.

So what does the mirror system contribute?

...[it] allows an individual to understand the action of others ‘from the inside’ p. 264

What does "from the inside" mean to R&S?

...the observed action is understood from the inside as a motor possibility and not just from the outside as a mere visual experience. p. 270

In other words, it is the "understanding" that I-can-do-that-too and nothing more.

Rizzolatti, G., & Sinigaglia, C. (2010). The functional role of the parieto-frontal mirror circuit: interpretations and misinterpretations Nature Reviews Neuroscience, 11 (4), 264-274 DOI: 10.1038/nrn2805


james said...

Hi Greg,

So I thought I would try and reply to your last post as I think you have slightly misinterpreted some the MN work.

As far as I am aware Rizzolatti and other have always thought that mirror neurons in F5c are encoding a feature of an observed grasp that is more than simply the grasp itself. This is because some the mirror neurons encode a grasp that seems to be effector non-specific. However, they also record mirror neurons that are strictly congruent. Indeed in their early papers they even hint at a hierarchical organisation of mirror neurons with F5c that encode the observed action at different levels of abstraction. In other words, one has to be careful when talking about "mirror neurons" as to whether you mean broadly tuned or strictly congruent. These are different populations.

Secondly, you are right when you say that when see someone pick up a cup you can not uniquely identify the intention. This is simply because the same action can be required for multiple intentions. This was pointed out in 2005 by Jacob and Jeanerrod in a TICs article. So any system that can infer an intention given an action, irrespective of whether this is done by mirror neurons, must be able to deal with this. The key point is not whether the system is able to infer the correct intention but rather can we infer the most likely intention given our past experience. In 2007 I published some papers that answered this using a predictive coding scheme. (Kilner et al. 2007 Neuroreport, Cognitive Processing).

You might me interested to note that it is precisely because of this one to many mapping problem that mirror neurons can not simply be due to sensory motor associations.


Greg Hickok said...

Hi James,

No I understand that there are more abstract mirror neurons, but if you rely on these cells to do the work, don't you lose the explanatory power of mirroring because it is no longer a motor mirror but an abstract goal/intention representation?

I don't understand how the one to many mapping problem argues against a sensory-motor explanation. Can you elaborate?

james said...

Hi Greg,

The one to many mapping is a problem only if you think that the motor firing in F5c is driven by the visual input - i.e is caused by sensory-motor associations as opposed to motor-sensory associations. The reason why is that the same visual act - grasping a nut say - does not uniquely map onto one motor output. We can perform the same action with very different muscle patterns. The converse is not true - given the motor output we can predict the visual consequences of that action. This is why it is quite common for recognition models to be predictive coding models, using generative rather than inverse models. This is what we showed in those 2007 papers.

The other related issue is how do sensory motor associations deal with situations where the observed action is occluded. In this situation the action is not seen so MNs can not be driven simply by associations between sensory motor areas. The system must be predictive.


Greg Hickok said...

Maybe I'm missing your point James, but two things come to mind:

1. The one-to-many problem is not unique to action perception. There's lots of ways I can reach for my cup too. We know that presentation of visual object stimuli activate "canonical neurons" presumably because the visual features of the object are relevant for guiding actions at various levels of planning. All I'm suggesting is that the same thing is happening for action observation.

2. No matter the direction of sensory-motor interaction (inverse or forward), this still doesn't imply that any meaning is encoded in the mapping.

james said...

Hi Greg,

No I do not think you have missed anything.

You are quite right that the one-to-many mapping problem is not unique to action perception. Indeed the original work was on lower level visual perception.

Second, you are also right that the predictive coding model does not imply that meaning is encoded. I merely wanted to point out that mirror neuron firing can not be explained by a simple sensory motor associative model. That is all.