Decoder co-adaptation · Grey Matter

Decoder co-adaptation is the closed loop in which the decoder of a brain-computer interface and the brain of its user learn at the same time, each adjusting to the other's errors until control becomes skilled.

Decoder co-adaptation. Decoder co-adaptation is the closed loop in which the decoder of a brain-computer interface and the brain of its user learn at the same time, each adjusting to the other's errors until control becomes skilled.

A neural decoder is first fitted to activity recorded while the user watches or imagines movements. Once the user drives the cursor, two things change. The person sees every error and corrects it, and the neurons in motor cortex shift their firing to suit the mapping they are given. When nine decoders were compared in monkeys in 2010, the assumptions that mattered most offline caused biases the animals simply compensated for in closed loop. ReFIT, from Stanford in 2012, used the loop from the machine's side: it retrained a Kalman filter on closed-loop data, with every velocity turned to point at the target, and halved the time two monkeys took to acquire targets; in 2015 the design gave two people with ALS more than double the performance of earlier pilot participants.

The brain can learn a fixed decoder. With the decoder held constant for days, monkeys' motor cortex settled into a stable activity pattern for the task, recalled quickly each day and resistant to interference, and a second one could be stored beside it.

Adapting the decoder speeds the start without spoiling the learning. Decoder updates improved early performance and absorbed changes in the recordings, while the neural side still improved, retained the skill and resisted interference from moving the real arm.

What the brain can learn quickly is limited. Mappings that used activity patterns the population already produced were learned within a session; patterns outside that repertoire mostly were not.

It is a laboratory for credit assignment. The experimenter chooses which neurons move the cursor, so it can be seen whether the brain finds and changes those cells in particular.

Control belongs to the pair, user and decoder together.

Each learner changes the problem the other is solving, so a decoder is designed for the loop it will run in.

Questions: When only some neurons drive a cursor, does the brain find and change those ones? Partly. In 2008 monkeys moved a cursor in three dimensions with a few dozen motor cortex neurons, and the experimenters then rotated the mapping for a subset of them, so those cells now pushed the cursor the wrong way. The animals adapted, and the changes in firing reflected two things at once: a general shift of strategy across the whole population, and adjustments weighted by how much each neuron contributed to the error. The brain can single out the cells that matter, at least in part, which is exactly the credit assignment problem solved on a small scale. Can the motor cortex learn any mapping a brain-computer interface gives it? Not within an afternoon. In 2014 monkeys controlling a cursor from motor cortex were given new mappings of two kinds: ones that could be driven by combinations of activity the recorded population already produced, and ones that required patterns it never showed. They learned the first kind readily within a session and largely failed at the second. The circuit's existing structure limits which new patterns are easy to make, which may be why new skills come faster when they resemble ones already learned. Why can the decoder that fits recorded data best lose once someone drives it? Because the user sees the cursor and corrects it. Offline, a decoder is scored on how well it reconstructs movements from activity that was recorded with no cursor to steer; in use, the person compensates for its errors as they happen. When nine decoders were compared in monkeys, the assumption that weighed most offline, about how the neurons' preferred directions are spread, produced directional biases the animals simply corrected for, and how the decoder smoothed the cursor became the largest difference. ReFIT drew the lesson from the other side: it was retrained on closed-loop data and halved the time to reach targets.