Monday, July 26, 2010

(Mis)understanding mirror neurons -- An alternative interpretation to "action understanding" and why they got it wrong in the first place

The idea that mirror neurons support action understanding is by far the dominant interpretation of the function of these cells in the monkey motor system. However, it is not the only interpretation. A "sensory-motor" hypothesis, such as that proposed by Cecelia Heyes and others, has been gaining steam in the last few years. In a just published piece in Current Biology, Marc Hauser and I propose a variant of the sensory-motor view, namely that mirror neurons function not for action understanding but for action selection, just like "canonical neurons" in the macaque motor system. We also outline why, in our view, mirror neurons were misunderstood from the beginning, namely that the more straightforward action selection interpretation was not supported by monkey behavioral data available at the time. We point out that the relevant behavioral data has now emerged from recent research and that the action understanding view requires serious re-evaluation. The outline of the argument is presented in the following summary from our paper:

It is hard to imagine a class of neurons that has generated more excitement than mirror neurons, cells discovered by Rizzolatti and colleagues in macaque area F5 that fire both during action execution and action observation. We suggest, however, that the interpretation of mirror neurons as supporting action understanding was a wrong turn at the start, and that a more appropriate interpretation was lying in wait with respect to sensorimotor learning. We make a number of arguments, as follows. Given their previous work, it would have been natural for Rizzolatti's group to interpret mirror neurons as involved in action selection rather than action understanding. They did not make this assumption because, at the time, the data suggested that monkey behavior did not support such an interpretation. Recent evidence shows that monkeys do, in fact, exhibit behaviors that support this alternative interpretation. Thus, the original basis for claiming that mirror neurons mediate action understanding is no longer compelling. There are independent arguments against the action understanding claim and in support of a sensorimotor learning origin for mirror neurons. Therefore, the action understanding theory of mirror neuron function requires serious reconsideration, if not abandonment. (p. 593).

Heyes, C. (2010). Where do mirror neurons come from? Neuroscience & Biobehavioral Reviews, 34 (4), 575-583 DOI: 10.1016/j.neubiorev.2009.11.007

Hickok G, & Hauser M (2010). (Mis)understanding mirror neurons. Current biology : CB, 20 (14) PMID: 20656198


The Vlad said...

An even simpler hypothesis might be that mirror neurons serve no function, but are simply an epiphenomenon of the way the monkeys are trained. I once asked Rizzolatti, via email, whether the experimenters ate food in front of the monkey. I don't recall the exact reply and can't drag up the email, but I do recall that he could not rule out the possibility. The reason I asked is that, should the experimenter happen to eat in front of the monkey during a time when the monkey itself was eating, then it remains possible that a neuron in motor cortex -- one that happens to receive the right sensory inputs and happens to control the right motor actions -- might learn to associate the two actions. This learning would have occurred incidentally, and no special significance need be attributed to the association (the association being a sensory-motor one, albeit it of a specific content). What evidence is there to specifically rule out this (null?) hypothesis?

Marc E. said...

I'm hoping you can elaborate on what you mean by 'selection' in the context of the mirror-neuron experiments. You list a number of cases where macaques do exhibit imitation, but in the experiments I've seen showing motor activation during perception, I don't recall seeing the macaques actually doing anything. Are you saying that they are 'thinking about' or 'considering' replicating the motor action they are seeing? Is there any direct evidence for that hypothesis (it seemed a lot of detail was left out probably due to length limits) aside from the citations which mainly argue against motor-understanding (5, 9, 10)?
Regardless, the idea that primates activate the motor system when conceptualizing doing motor movements does seem interesting in and of itself and not too far removed from the idea that understanding may be facilitated by simulation.

Greg Hickok said...

Hi Vlad,
You make an important point. In fact, I believe it highly likely that the training procedure itself is critically important in tuning the response properties of mirror neurons. This is predicted if in fact mirror neurons are sensory-motor association cells that can make arbitrary associations between perceived actions are appropriate responses by the animal. What was lacking was a clear behavioral demonstration that actions (and not just objects) can serve as the input to the motor execution system. On this view, mirror responses are just a (perhaps minor) subtype of action-action associations.

Greg Hickok said...

Hi Marc E,

What we are saying is that mirror neurons function exactly as "canonical neurons" do. They take action-relevant sensory input -- objects for canonical neurons and actions for "mirror neurons" -- and associate them with appropriate response actions.

If you look at canonical neurons you find that they too respond simply upon presentation of an object, even without the monkey generating an action. They also respond when the monkey reaches for that action. I.e., they are sensory-motor cells that seem to be involved in performing the mapping between sensory information and motor actions. Mirror neurons, our claim is, are identical.

Think of it this way: I present you with a cup of coffee (a visual object) and this activates cells involved in generating an appropriate grasping response toward the cup. These are canonical object-oriented cells. Now suppose there is a cup of coffee in front of the two of us and we both want it. You see me starting to reach for it. My action activates your cells involved in coding a competitive reaching action. These are mirror neurons. These circuits likely work in parallel.

Notice that not all action-triggered actions have to mirror. If you see me holding a cup of coffee and I suddenly start a motion that will result in me throwing the liquid in your face, the appropriate response is not a mirror gesture, but a ducking or dodging gesture. This is what I mean when I say that mirror cells are likely to be a subset of action-action association cells.

I hope this helps clarify...

Antonio said...

Dear Prof. Hickok

I follow with great interest your discussion on the true meaning of mirror neurons. I read your recent, and deep, contributions to this discussion published on Current Biology and Language and Cognitive Processes. In both is not mentioned the study of Mukamel et al. that shows the existence of mirror neurons in humans. Probably, Mukamel et al's paper was published after you had already ended writing both your contributions. I would be interested to know what you think about this finding, and especially about the fact that mirror neurons were recorded in cortical areas that are not part of classical mirror system.



James Kilner said...

Hi Greg,

I thought this correspondence to Current Biology was really interesting. I just had a quick thought though. In the two route model VM associations are presumably learned in both pathways. In other words the semantic information and how it influences the motor set must be learned in a similar manner to the VM associations in the dorsal pathway. The most likely mechanism of both sets of learning is associative learning. If you agree that this is the case couldn't the counter mirroring studies you cite just as easily be modulating the ventral and not the dorsal pathway? In other words when learning do people just learn a semantic association between the observed and executed actions and not the more pragmatic associations?


Greig de Zubicaray said...


The NeuroCritic has covered the Mukamel et al. Current Biology report and has some insightful commentary:



Greg Hickok said...

Hi James,

Sorry for the slow response...

The dorsal and ventral systems are good for different things. The dorsal stream is good for associating sensory features such as shape, size, location, and orientation with the motor system, which is critical for accurate actions directed toward an object. You don't need understanding to do this. The ventral stream is good for mapping sensory features onto stored semantic information, e.g., that the thing you are looking at is a cup, independent of its particular shape, size, and orientation. The goals of these two pathways are to some extent at contrary purposes: for say grasping you want to code the details of the particular shape, size, and orientation of a cup such that a large upright mug will be treated differently by the system than a small mug on its side. But if you are trying to figure out *what* an object is rather than *how* to grasp it, you need to ignore (discard) all these low-level featural differences and extract the invariant object properties.

So now to your question: yes, I would imagine that we can form associations with the motor system in both the dorsal and ventral streams. But only in the dorsal stream will these associations actually be useful for grasping because this stream codes the lower-level featural details. The ventral stream will be pretty useless for grasping because all it will tell you is that you are looking at a CUP. You may know from this that you desire to grasp it, but its location, size, etc. will not be specified. What this means is that the input to the motor system from the ventral stream will be fairly high level (this is an object I want to reach for) whereas the input to the motor system from the dorsal stream will be lower level (here is how to actually grasp it).

I hope this helps...