Feedforward Transformations and Internal Models*

We can, and do, use feedback from vision and other sense to guide movements. For example, one could visually compare hand position and goal position and keep moving the hand until one sees that they align. In engineering, this is known as ‘Servo Control’. Servo control has the advantage of being very accurate and robust, i.e., even if the control system has some internal errors: as long as the feedback drives the effector in approximate the right direction, it will get there. However, it requires rapid feedback.

In biological systems neural transmission has finite delays. For example, visual feedback to the cerebral cortex occurs in the order of 100ms. By that time, many saccades have already finished, and a rapid arm movement is well underway. Even after that, if one relied on slow visual feedback for a fast movement, the hand would tend to oscillate toward toward the goal as it overshoots, corrects, overshoots, etc.

The alternative is to use feedforward mechanisms that generate movement commands based on initial external and internal conditions. This does not preclude the use of rapid internal feedback loops that make use of efference copies / corollary discharge. When we say feedback in this context, we mean sensory feedback, unless otherwise stated.

Most investigators think that rapid goal-directed gaze shifts are entirely ballistic, governed by feedforward mechanisms (but see Klier and Crawford 1998), and goal directed reach movements are governed by a combination of feedforward and feedback mechanisms. Here, we will focus on the feedforward mechanisms common between both.

Whereas the advantage of feedforward is speed and prediction, the challenge is that the commands generated must be very accurate. To do this, the transformation must possess an accurate internal model of the the source of any sensory information, the effector being control, and importantly for our topic, the spatial relations between them. Moreover, these internal models must be well calibrated through practice. This is what separates Olympic athletes, ordinary people, and some patient populations.

Many motor control investigators have found it useful to distinguish between ‘Forward’ and ‘Inverse’ internal models. Forward models convert an internal command into a predicted state (e.g., final hand position), that is, in the same order it would occur. Inverse models convert a desired state into the required command to achieve that state.

These different sets of terminology can often be confusing because inverse models are generally used in the serial (feedforward) generation of the command, whereas forward models are often seen within internal feedback loops. For example, to achieve a desired position, one might put this signal through an inverse model, and then put the output command both to the effector (to generate movement) and to a forward model (to generate a predicted position that could subtracted from the initial desired position to create an internal feedback loop).

Inverse and forward models can be seen within formal models of both the oculomotor and reach systems, but this terminology is used less frequently in the former – perhaps because oculomotor models were better established when it came into vogue.

* This section will likely come earlier in the review paper, and we may not need to talk about inverse and forward models. I just wanted to clear up the difference between this stuff and feedforward / feedback terminology.