Brain-Computer Interfaces Flashcards
Define neural engineering and give two examples
Electrical interventions in the nervous system and intervenes at the level of the action potential.
Cochlear Implants and Retinal Prosthetics are two examples.
What is a way that deep brain stimulation can be use?
It cam be implanted to stimulate the basal ganglia and treat dystonia and Parkinson’s Disease.
What is dystonia and how does DBS help?
Dystonia is uncontrollable shaking and difficulty controlling extremities. DBS to the basal ganglia makes limbs more controllable and greatly reduces shaking.
DBS
Deep brain stimulation
BCI
Brain-Computer Interface
Degrees of freedom problem in motor control
Also known as the motor equivalence problem, this is the problem of there being many ways for motor control to achieve a goal.
Why is reaching difficult?
It takes years to learn.
Redundant degrees of freedom.
Muscles are complex – the same input can yield different responses depending on the muscle’s position, velocity, and force.
Spinal reflexes and hierarchical control (there are many aspects working in parallel)
Inputs and outputs are in different formats (visual-motor)
Preplanning is necessary
Why is preplanning motor movement necessary?
Relying on feedback is too slow.
What types of factors go into hand movement?
Join angles, Joint velocity, Join Acceleration, Hand velocity
Define velocity
The velocity of an object is the rate of change of its position with respect to a frame of reference, and is a function of time.
How does reach differ in joint coordinates versus visual coordinates
In visual coordinates the hand moves straight, but it’s more complex in joint coordinates.
What is Fitts’s Law?
A predictive model of human movement primarily used in human–computer interaction and ergonomics. This scientific law predicts that the time required to rapidly move to a target area is a function of the ratio between the distance to the target and the width of the target.
ie. A smaller target farther away is harder to reach.
What is the 2/3 power law?
Speed is related to curvature. Is use in talking about handwriting. Links curvature trajectory to angular velocity.
Sounds to me like it’s saying that 2/3 power law says that handwriting is related to the speed of the curve of the hand to the velocity of the angular direction in relation to that.
2/3 is an exponent
Scale and effector invariance
These do not change with different body parts. My toes have the same hallmarks as my hands for writing.
effector: an organ or cell that acts in response to a stimulus. (So these do not change if I’m writing with my toes???)
Both are multiplied by a common factor to get the invariance
Define effector
an organ or cell that acts in response to a stimulus.
What are some regularities of movement?
Fitts's Law 2/3 power law scale and effector invariance Repeated movements are very similar Movements are smooth
Three theories of motor contorl
Signal-Dependent Noise (Wolpert and Harris) Optimal Feedback Control (Todorov and Jordan, and Steve Scott) Internal Models (Masao Ito and many others)
What is signal-dependent noise?
Signal-dependent noise determines motor planning
For a given movement duration, the neural command minimizes error or for a given error tolerance, the neural command maximize speed
What is a poisson distribution
is a discrete probability distribution that expresses the probability of a given number of events occurring in a fixed interval of time and/or space if these events occur with a known constant rate and independently of the time since the last event.
Think the mail example
What are some safe assumptions of signal-dependent noise theory?
You move as quickly, smoothly, and accurately as noise allows.
The goal is to make accurate movements
Neural control signals are Poisson: the noise scales with the signal
Noise accumulates over the duration of the movement
Neurons are also variable in that particular details will change about individual neuron firing each time.
Can optimal feedback control theory be falsified?
No
What does Optimal Feedback Control theory say?
There must be redundancies because there is multiple input and output.
Redundant noise can be exploited.
You don’t correct error in task performance that don’t hurt you. Also considered a minimal intervention principle for this reason.
What are Internal Models?
They are models where the brain creates an internal representation of the body and the environment and uses that to plan movements and to anticipate consequences of actions.
Most of the support for this is in behavioral rather than physiological evidence.
Another way to look at this from wikipedia:
It is a process that simulates the response of the system in order to estimate the outcome of a system disturbance.
The internal model theory of motor control argues that the motor system is controlled by the constant interactions of the “plant” and the “controller.” The plant is the body part being controlled, while the internal model itself is considered part of the controller. Information from the controller, such as information from the central nervous system (CNS), feedback information, and the efference copy, is sent to the plant which moves accordingly.
Within the Internal Model, what is the inverse model?
Inverse models use the desired and actual position of the body as inputs to estimate the necessary motor commands which would transform the current position into the desired one. For example, in an arm reaching task, the desired position (or a trajectory of consecutive positions) of the arm is input into the postulated inverse model, and the inverse model generates the motor commands needed to control the arm and bring it into this desired configuration. They are a type of internal model.
What is efference copy?
an internal copy of an outflowing (efferent), movement-producing signal generated by the motor system
What are forward models?
In their simplest form, forward models take the input of a motor command to the “plant” and output a predicted position of the body. They are a type of internal model.
What is a plant?
A body part.
What are some non-ideal aspects of muscles?
Force depends on length and velocity.
They are sluggish.
They only work in one direction (antagonist arrangement)
Antagonist arrangement
a muscle whose action counteracts that of another specified muscle.
Antagonist and agonist muscles often occur in pairs, called antagonistic pairs. As one muscle contracts, the other relaxes. An example of an antagonistic pair is the biceps and triceps; to contract - the triceps relaxes while the biceps contracts to lift the arm.
What is some evidence that M1 codes muscles?
A monkey is trained to rotate a disc to a certain point. Then a weight is added to the disc to make use of muscle extensor or flexor more necessary. This means that the movement is the same, but the engagement of the muscles is different. Neurons increased their rate of firing in response to additional weight.
A. Direction-specific increases in single unit activity were observed when there was no load applied to the hand during flexion-extension movements of the wrist. A particular single unit increased its activity for either a flexion or extension movement and was either reduced or quiescent during movement of the opposite direction. This observation alone is consistent with both the “muscle dynamics” and “displacement” hypotheses.
B. Regardless of movement direction, the absolute firing rate for a given single unit was variable but predictable given the direction, magnitude, and rate of change of force opposing wrist movement. This observation is consistent only with the muscle dynamics hypothesis.
What is one thing that muscle spindles do?
Regulate posture. Seen in myotactic reflext.
Myotactic Reflex
The stretch reflex (myotatic reflex) is a muscle contraction in response to stretching within the muscle. It is a monosynaptic reflex which provides automatic regulation of skeletal muscle length. When a muscle lengthens, the muscle spindle is stretched and its nerve activity increases.
What is extensor?
Muscle extension.
PTN
Pyramidal Tract Neurons
What is evidence that M1 codes movements?
Center-Out reach task. Movement is predicted by population vectors.
What is the center-out reach task?
The subject has to begin with their hand on something in the center, they receive a target to move to but then must wait for a cue before moving. They get the cue and move their arm to a target outside of the center. This has shown planning, firing of neurons before the beginning of movement.
What is evidence that M1 codes for movements and muscles?
The study that determined this looked at extrinsic motion (direction involved in moving the wrist), muscle use (muscles involved in moving the wrist), and joint use (direction of joint change).
The monkey held its hand in a pro, sub, or mid position. Then he flexed the wrist in 8 different directions according to a cue on the screen. Also considered was the preferred direction of the muscles involved in that movement and how those preferred directions were reflected in different wrist starting positions.
Point being, all of the variables could be measured independently and it was found that they all involve activity in M1. This activity isn’t on the same neuron though.
Intrinsic movement
This is the type of movement. In other words, flexing my wrist is the same type of movement regardless of my wrist starting position.
Extrinsic movement
This is how the body moves through space. In other words, flexing my wrist involves different movements through space depending on my starting point. I can flex and move down or flex and move up.
What are three types of neurons identified in the Kakei, Hoffman, Strick paper?
Muscle-like, Extrinsic-like, Extrinsic-like but modulated by posture. These show that M1 codes for muscle and movement.
What does PD stand for?
Preferred direction
What does the Kakei, Hoffman, Strick paper say about the preferred direction of neurons?
The shifts of neurons parallel those of muscles. (I really don’t get this too much).
What is the Reference Frame paradigm?
Refers to eye-hand coordination, written here as the eye-to-hand reference frame. Includes visual receptive fields combined with proprioceptive receptive fields. We have to coordinate what we see with where our body is in space.
In his words: Visual receptive fields combine with proprioception to yield limb-centered tuning which drives muscles.
So goal directed reaching requires a reference frame transformation.
What is tuning?
Neuronal tuning refers to the property of brain cells by which they selectively represent a particular type of sensory, association, motor, or cognitive information.
What is the hierarchical path for the reference frame transformation?
Retina LGN V1 PPC (posterior parietal cortex) PMd (premotor dorsal area) M1 Spine Muscles and then back again
This is a feed forward model, taking in the mean firing rate of single neurons (mean because neurons are variable?).
What is PPC?
Posterior parietal cortex
Problems with M1 codes dynamics?
Doesn’t show much that can’t be shown through SNR and you have to throw away the mean response. To see the variability you have to take out the mean.
What does it mean that M1 codes dynamics?
When I reach for something my path changes and neural firing changes. There’s a lot of variability but perhaps a dynamic pattern is visible. For instance variability decreases on stimulus onset.
This may allow viewing patterns that escape single neuron recording.
What is consistent across neurons (but step one is throw out the mean?)
What are BCI algorithms attempting?
To replace the effector
What are some parts involved in using a BCI?
Calibration: Scientists move the cursor, record neural activity, and then map what happens in the brain to cursor kinematics.
(Observation-based calibration)
Control the BCI:
Generate the mapped patterns to control the cursor
What is kinematics?
the branch of mechanics concerned with the motion of objects without reference to the forces that cause the motion.
the features or properties of motion in an object.
What are the three decode algorithms?
Population vector
Linear Filter
Kalman filter
What is involved mathematically in population vector algorithms?
- Fit each neuron with a cosine (gets to vote)
- Estimate the preferred direct (add up votes)
- Weight the preferred direction by the cell’s firing rate.
- Sum all the weighted vectors.
What can be used to calculate population vectors?
Center-Out Reach task
What is a linear filter?
Each spike contributes a little pulse of movement and all of those need to be added together.
There’s a training matrix algorithm and a testing (decoding) one. I doubt I’ll need to know these but they are on the slide.
Uses linear algebra
Requires finding lots of parameters
What is the Kalman filter?
It was invented to track satellites is a continuous Bayes estimator.
It has a state equation and an observation equation.
It offers uncertainty weighting by adding in noise
Also uses known statistics of hand movements.
Combines a trajectory model and observation model that can compute each timepoint.
What are the known statistics of hand movements?
arm state x^t=[Position, velocity, acceleration]
Determined in the center-out reach task
Used in Kalman filter
What is a problem with BCI?
Calibration is hard to escape. One monkey’s neural pattern doesn’t apply to another monkey or to the same monkey the next day.
Which algorithm is best?
Depends on what you think M1 does.
All of the algorithms decode velocity.
Many new algorithms are coming out all the time.
Plasticity probably erases the differences between them.
Amount of experience and implementation details like parameter numbers and quality of recording can skew results. (how you use them matters?)