Jose Carmena,

Studies robotic monkeys.

From lands far away.

His call to monkeys

“Brain-machine interfaces

Let you move again”

BMIs decode

Neural spikes, feed back to brain,

And create circuits.

“But while brains adapt,”

Jose and monkeys wondered,

“Can the BMI?

And if it did change,

Could it impair performance,

As others suggest?”

But CLDA* –

To new circuits it adapts,

Parameters change.

Monkeys had a task –

Control the cursor with brain,

And reach the target.

fig1_taskSetup

Fig 1. Experimental setup, and CLDA. A, B – CLDA used on day 1 to improve performance. Decoder then kept fixed. C, D – Two monkeys trained to hold at center, then reach for target. Monkeys used an arm or BMI to control a cursor.

CLDA used

Until monkey had success.

Then, only brain changed.

Improved performance,

Substantial by Wilcoxon,

Accurate movements.

behavioralComposit_v2

Fig 2. Behavioral performance in two-learner system (i.e. with neural and decoder adaptation). A,C – Performance across time in both monkeys improved. B – Reach trajectories across time. D – Average improvements in performance; tested by one-sided, paired Wilcoxon sign-rank test. “Seed” refers to performance with initial decoder. “Early” refers to day 1 performance. “Late” refers to best performance after day 1.

Neuron tuning curves,

Sharpened across training days,

Looked like performance.

So, CLDA

Stabilized cortical maps.

Effect was success.

fig3_v3_tuningChanges

Fig 3. Changes in neural tuning. A, B, C – Fitted tuning curves, across time, for various sample neurons. D – Changes in MD and PD across time; grey indicates neurons not significantly tuned. E – Pairwise correlation of ensemble tuning maps across time. F – Map correlations for each day (red) overlaid on task performance (black).

“But two things adapt,”

Jose and monkeys wondered,

“Which matters the most?”

With CLDA,

More improvement was observed.

Mirrored tuning change.

MD** and PD***,

Neuron tuning properties,

Varied with success.

If BMI wrong,

Assigned wrong MD/PD,

Tuning tried to match.

Once learning is done,

Neurons reach peak earlier.

Initiate reach.

Neurons can adapt

To the decoder employed

In tuning and time.

fig4_v4_tuningChangeMechanisms

Fig 4. Extent of neural adaptation is related to performance improvement. A – Increases in performance, and correlation between tuning and performance, for three samples, one of which had no CLDA applied (black trace). B – Map correlation on day 1 against change in task success. C – Ensemble-averaged change in MD or PD against change in task success. D – Fraction of BMI units with significant changes in MD and PD. E, F – Change in MD or PD against decoder weight (E) or error (F) assigned to each unit for either MD (E) or PD (F). Black lines are linear regressions.  G – Time histogram for sample units. H – Onset of direction tuning (left) and peak firing rate (right), early or late in learning.

“Still, one last question –

New circuits resist old ones?”

Said Jose et al.

“Because old circuits

Can overlap with new ones,

Interfere in task.”

Monkey maintained force,

While it performed the reach task,

At interspersed times.

This impaired success,

Disrupted monkey learning,

Only in SC****.

Impairment profound.

Disrupted neural tuning.

Lessened by learning.

fig5_simultaneousControl

Fig 5. Native arm movements do not interfere. A, B – Monkey J performed BMI-SC (simultaneous control) task. J first acquired force target with arm, then used BMI to move cursor. C – Performance across time. Blue dots are when CLDA was applied. D – Average BMI and BMI-SC task performance. E, G – Comparison of MD and PD in both tasks. F – Fitted tuning curves for two sample units.

Jose and monkeys.

CLDA and brain change.

They improve success.

They look to future.

Optimization is next.

More robot monkeys.

————————————————————–

Notes:

* Closed loop decoder adaptation

** Modulation depth

*** Preferred direction

**** Simultaneous control of both force and BMI-controlled reach task

Open Archive article:

Orsborn et al. (2014) Closed-Loop Decoder Adaptation Shapes Neural Plasticity for Skillful Neuroprosthetic Control. Neuron, p. 1380 – 1393.

____________________________________________________________________________________________

Javier How is a 1st year Neuroscience student currently rotating with Saket Navlakha at the Salk Institute. Upon a second reading of this post, he remembers why he’s not in an MFA program.

Advertisements

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s