Thursday, October 7, 2010

Internal forward models -- New insight or just hype?

In case you haven't noticed, the concept of internal forward models -- an internal prediction about a future event or state -- are all the rage. The concept comes out of the motor control literature where one can find pretty solid evidence that motor control makes use of forward predictions of the sensory consequences of motor commands (e.g., check out the seminal paper by Wolpert, Ghahramani, & Jordan, 1995). These concepts have been extended to speech (e.g., Tourville et al. 2008; van Wassenhove et al., 2005) and there has been a ton of work trying to establish the neural correlates of these networks (e.g., see Golfinopoulos et al. 2009; Shadmehr & Krakauer, 2008), recent work suggesting an association with clinical conditions such as aspects of schizophrenia (Heinks-Maldonado, et al. 2007) and stuttering (Max et al. 2004), and even applications of the concept of high-level cognition such as "thought" (Ito, 2008), as well as applications to social cognition (Wolpert et al. 2003) with links to the mirror system (Miall, 2003).

I'm a big fan of control theory in general and I think there is a lot to be gained by thinking about speech processes in these terms. At the same time, I'm a little uncomfortable with the widespread application of these models. It kind of reminds me of the mirror neuron situation in that a framework for thinking about one problem is generalized to all kinds of situations. I'm also a bit uncomfortable about the assumed tethering between forward models and the motor system. A forward model is just a prediction. In the context of motor control, it makes sense to make predictions (e.g., sensory predictions) based on the likely outcomes of motor commands. But more generally, predictions can come from lots of sources. Perceptual fill-in processes are a kind of forward model: the visual system for example makes predictions about the color and texture of a given portion of the visual scene based on the color and texture around that region. One can predict the consequences of an ocean wave hitting a rock based on past perceptual experiences. So forward models don't have to come from the motor system and there are probably lots of systems and mechanisms that generate predictions (forward models). It is worth having a look at Karniel's (2002) short comment, "Three creatures named 'forward model'" for some cautionary discussion.

So is the internal forward model concept just hype? No, I don't think so. It has already demonstrated its utility in the motor control literature and there are systems in the brain that appear to support motor-related forward models (cerebellum is one, posterior parietal cortex is another). There are some real insights to be gained from this framework in the speech domain as well, but I think there is the danger of over-application of the concept and we need to proceed cautiously.


Golfinopoulos, E., Tourville, J.A., and Guenther, F.H. (2009). The integration of large-scale neural network modeling and functional brain imaging in speech motor control. Neuroimage 52, 862-874.

Heinks-Maldonado, T.H., Mathalon, D.H., Houde, J.F., Gray, M., Faustman, W.O., and Ford, J.M. (2007). Relationship of imprecise corollary discharge in schizophrenia to auditory hallucinations. Arch Gen Psychiatry 64, 286-296.

Ito, M. (2008). Control of mental activities by internal models in the cerebellum. Nat Rev Neurosci 9, 304-313.

Karniel, A. (2002). Three creatures named 'forward model'. Neural Networks 15, 305-307.

Max, L., Guenther, F.H., Gracco, V.L., Ghosh, S.S., and Wallace, M.E. (2004). Unstable or insufficiently activated internal models and feedback-biased motor control as sournces of dysfluency: A theoretical model of stuttering. Contemporary Issue in Communication Science and Disorders 31, 105-122.

Miall, R.C. (2003). Connecting mirror neurons and forward models. Neuroreport 14, 2135-2137.

Shadmehr, R., and Krakauer, J.W. (2008). A computational neuroanatomy for motor control. Exp Brain Res 185, 359-381.

Tourville, J.A., Reilly, K.J., and Guenther, F.H. (2008). Neural mechanisms underlying auditory feedback control of speech. Neuroimage 39, 1429-1443.

van Wassenhove, V., Grant, K.W., and Poeppel, D. (2005). Visual speech speeds up the neural processing of auditory speech. Proc Natl Acad Sci U S A 102, 1181-1186.

Wolpert, D., Ghahramani, Z., & Jordan, M. (1995). An internal model for sensorimotor integration Science, 269 (5232), 1880-1882 DOI: 10.1126/science.7569931

Wolpert, D.M., Doya, K., and Kawato, M. (2003). A unifying computational framework for motor control and social interaction. Philos Trans R Soc Lond B Biol Sci 358, 593-602.


Anonymous said...


The nice thing about forward models is that they are probably the primary source of information to drive movements. The motor system adjusts the motor commands only when the actual and predicted sensory consequences of the movement differs. Therefore, it is nothing like using sensory input to drive the behavior. Sensory input are only there to signal when the output of the forward model was wrong.

I think that there is plenty of evidence to support the cerebellum as the locus of forward models (TMS: Miall et al. 2007 PLOS biology; patients: Smith et al. JNP 2005 or Criscimagna-Hemminger et al. 2010). Even for internal models of moving objects have been found in the cerebellum (Cerminara et al. J Physiol 2009).

I have to say that I'm very skeptical about the PPC as site for forward models. The study by the team of Richard Andersen (in PNAS)is far from convincing. The task they used does not discard the possibility that the observed neural activity is not driven by very short-term prediction of the moving cursor rather than prediction of the hand movements.

The real question is if you are able to predict someone else's movement or intention, have you built a forward model of that person? Would such forward model be located in the cerebellum as well?


PS: I enjoyed reading your piece in JOCN about the problems of the mirror neuron theory

Anonymous said...

Great entry. Personally, I think that the evidence (to date) for motor-system instantiated internal forward models being a necessary component of speech perception is unconvincing.