Eeny, Meeny, Miny, Moe...

One day, I decided to make muffins. Unable to find a hot-pad when they were done, I grabbed a towel to protect my hand from the hot pan. I misjudged the thickness of the towel and quickly found my hand a bit too warm for comfort. My first reaction was "HOT HOT HOT! Drop it now!" followed by "No! Don't drop the muffins!" I had a few options: drop the pan and save my hand but lose the muffins, save the muffins and burn my hand, or try to find some sort of compromise that would save both. Rather hungry, I instinctively went for the latter. Unfortunate for me, my brain's best solution was to switch the pan to my other, completely unprotected, hand, giving me a double scalding. At least I had some tasty muffins...

Our nervous system is constantly confronted with these sorts of choices. Granted, the above anecdote is rather complicated by motivational issues and more complex decision making processes. But we regularly choose between possible movements. Even if we know where we want to move, we still have to decide how to execute that movement since our limbs have an incredible number of degrees of freedom. How do we narrow down the infinite possible movements to just one?

Paul Cisek and John Kalaska’s 2005 Neuron paper (“Neural Correlates of Reaching Decisions in Dorsal Premotor Cortex: Specification of Multiple Direction Choices and Final Selection of Action”) studied the motor processing stream to examine how the brain plans movements when presented with 2 possible options. They found that the motor planning-execution stream acts as a funnel of sorts. Premotor dorsal cortex (PMd, commonly thought of as a ‘higher-level’ motor planning area of the brain) appears to plan movements to all potential targets, but once a decision is made, only information for the executed reach is passed on to the primary motor cortex (M1) to be implemented.

Cisek and Kalaska compared PMd and M1 activity when a monkey was presented with one or two potential reaching targets. In the one-target task, the monkey was briefly shown a single target, whose location he had to remember during a short delay period. A colored disk (appearing at the center of the screen) cued the monkey to make a reach to this remembered target location. In the two-target task, the monkey saw two differently colored potential reach targets. The color of the go cue told him which to choose. The researchers then examined the activity of neurons during these different task epochs.

This study focused on one particular property of motor cortex neurons known as tuning. Many neurons in the motor cortex seem to fire a lot when an animal makes or plans a movement in one direction (say right), and very little for the opposite direction (left). The firing for movement directions in between falls off as you get further from the neuron’s ‘preferred’ direction. Tuning describes these direction-preferences. Neurons in the motor cortex have preferred directions that roughly span the full movement space, and the so-called ‘depth of modulation’ (i.e. how strongly tuned it is) can vary from neuron to neuron. Also, these direction preferences may only appear during certain points in the movement planning-execution process.

Cisek and Kalaska recorded from several hundred neurons in PMd and ~40 neurons in M1 to compute their tuning functions as the monkey performed both tasks. Interestingly, this large collection of cells clustered into 4 categories, which seem to represent different stages of movement planning.

Potential response (PR) neurons: During the one-target task, these neurons showed tuning as soon as the movement target was presented. That is, they fired a lot when the targets were presented near their preferred direction, and this selective firing started as soon as the target was shown. The firing lasted through the delayed memory period until the movement was executed. But more interestingly, during the two-target task, PR neurons showed this tuning to both potential movement targets. Once the final movement target was cued, the activity shifted to represent only this location. PR neurons seem to help plan all potential reaches. Fittingly, these neurons represented nearly half of the recorded neurons in PMd and were only rarely found in M1.

Selected response (SR) neurons: Next-in-line in on the movement planning stream, SR cells only showed tuning after information about the final reach target was shown to the monkey (i.e. after the single target cue in a one-target task, or after the correct choice was cued in the two-target task). That is, they represented, and perhaps help plan, reaches to the selected movement direction. Again, these planning-oriented neurons were found primarily in PMd.

Build-Up (BU) and Movement (M) neurons: These cells both seemed linked to executing the planned reach. BU cells slowly developed a tuning for the final movement direction once the color cue was shown, gradually ramping up their activity as the monkey prepared to reach to the decided target. M neurons, however, only showed tuning after the go cue (i.e. once the monkey knew where he was reaching and started the movement). These cell types made up the majority of M1 neurons, and are not particularly new—several studies before this have found these sorts of movement-tuned neurons.

The four cell types Cisek and Kalaska found make a nice schematic picture of movement planning. We represent each potential movement target until we decide. After some choice is made, only the final movement information is passed down-stream. Finally, execution-related cells gear up to make the winning movement.

Movement planning results like this have interesting applications for brain-machine interfaces. There are two schools of BMI techniques: one based on decoding movement execution-related neural activity, and another focused on detecting this sort of higher-level planning information. If we can detect where someone is planning to reach, then we can drive a robot/cursor in that direction. Krishna Shenoy at Stanford, for example, has had some amazing success with this approach. These sorts of signals, however, don’t really allow for online corrections. Online feedback plays a huge role in our dexterous movements. Without it, we tend to get a bit inaccurate—think about making a really fast movement, where you can’t correct it, versus slowly reaching. Though less suited for motor BMI, this sort of ‘discrete’ BMI does lend itself incredibly well to other types of decoding, such as speech neuroprostheses. But, I think planning information is still potentially very useful for motor applications. For instance, using PMd activity to act as a prior, in your execution decoding could help reduce errors.

This study suggests that we plan several movements at once until we've decided, a very intuitive idea. It surely helps reduce our reaction times, and explains some common motor goofs. Ever done some weird combination of two movements when you had trouble deciding how to do something? Or said a mish-mash of two words you'd considered using?

But one huge lingering question is how motor cortex cells pick the ‘winning’ movement. How does the information transition between PR and SR neurons? This is where things surely get messy, with inputs from higher cognitive areas related to motivation, goals, visual processing, and so on all come in to play. I’m not signing up to try to decipher that sort of information-soup, but I’m glad others are on the case.


[Apologies for the long delay between posts. I caught a ridiculous head cold that knocked me out of commission for 3 weeks. I'm only now catching up.]