This course provides an introduction to basic computational methods for understanding what nervous systems do and for determining how they function. We will explore the computational principles governing various aspects of vision, sensory-motor control, learning, and memory. Specific topics that will be covered include representation of information by spiking neurons, processing of information in neural networks, and algorithms for adaptation and learning. We will make use of Matlab/Octave/Python demonstrations and exercises to gain a deeper understanding of concepts and methods introduced in the course. The course is primarily aimed at third- or fourth-year undergraduates and beginning graduate students, as well as professionals and distance learners interested in learning how the brain processes information.
Eb Fetz on Bidirectional Brain-Computer Interfaces
Loading...
Skills You'll Learn
Computational Neuroscience, Artificial Neural Network, Reinforcement Learning, Biological Neuron Model
Reviews
4.7 (513 ratings)
- 5 stars374 ratings
- 4 stars108 ratings
- 3 stars24 ratings
- 2 stars5 ratings
- 1 star2 ratings
JB
May 25, 2019
I really enjoyed this course and think that there was a good variety of material that allowed people of many different backgrounds to take at least one thing away from this.
RC
Mar 03, 2019
Great course! Really enjoyed the variety of topics and the just enough computational work in the quiz's. And that Eigen hat had me smiling and laughing about it for a week.
From the lesson
Learning from Supervision and Rewards (Rajesh Rao)
In this last module, we explore supervised learning and reinforcement learning. The first lecture introduces you to supervised learning with the help of famous faces from politics and Bollywood, casts neurons as classifiers, and gives you a taste of that bedrock of supervised learning, backpropagation, with whose help you will learn to back a truck into a loading dock.The second and third lectures focus on reinforcement learning. The second lecture will teach you how to predict rewards à la Pavlov's dog and will explore the connection to that important reward-related chemical in our brains: dopamine. In the third lecture, we will learn how to select the best actions for maximizing rewards, and examine a possible neural implementation of our computational model in the brain region known as the basal ganglia. The grand finale: flying a helicopter using reinforcement learning!
Taught By
Rajesh P. N. Rao
Professor
Adrienne Fairhall
Associate Professor
[MUSIC]. It is a great pleasure to introduce my
colleague and friend, Dr Eb Fetz, who's a professor of Physiology and Biophysics
here at the University of Washington. Eb got his Ph.D in Physics from MIT in
1967. He's widely regarded as one of the
leading researchers in the area of modern neuroscience.
Eb is also a founding father of the field of brain computer interfacing.
In 1969, he showed that monkeys can control the activity of single brain
cells to move the needle of a meter to get food rewards.
More recently he has proposed the concept of bidirectional brain-computer
interfaces, which can both record from and stimulate neurons at the same time.
To tell us more about this exciting new area of research, here's Eb.
>> Okay. Thank you very much.
so I'm going to tell you about Bidirectional brain-computer interfaces.
This is a new experimental paradigm that's illustrated schematically here.
And basically it consists of a computer that's connected to electrodes that
record activity from the brain. And process it in real time, and deliver
activity dependent stimulation back to the brain.
So this in a sense creates an artificial recurrent connection that operates
continuously during free behavior. And the brain can learn to incorporate it
into its normal activity. It also creates conditions that promote
synaptic plasticity. These are themes that you've heard about
in this course, and we'll take a look at some of these issues as we've explored
them recently in the lab. So everyone is familiar with the standard
brain computer interface, in which neural activity is record from let's say cells
in the brain. And processed to control an external
device such as a robotic arm. And the best cells for this sort of
function are in the motor cortex, actually.
Where neurons are used to controlling the activity of peripheral muscles through
their connections to the spinal cord. The ideal recording is from neurons, but,
as you've heard. They're also possibilities of recording
not only in neural activity from a single neurons with invasive intracortical
electrodes. But also, less invasively signals from
the surface of the brain, the electrocorticogram.
Or a surface of the scalp, even the electroencephalogram, EEG.
And these signals are processed then through electronic circuitry to generate
control signals for the external device. It could be cursor on a screen, or it
could be the movement of a prosthetic arm.
And the subject sees this output and uses this visual feedback to optimize the
control of these signals by modifying illustional controls of neural activity.
So that's a standard brain computer interface.
the, bidirectional, computer interface converts this activity to control a
stimulator. Which can electrically stimulate the
periphal muscles or similarly could be delivered in spinal cord or they could
actually be delivered anywhere in the brain.
So this creates a recurrent loop that can as I said operate continuously.
And this component of this, of the loop can be implemented in electronic
circuitry that can be made small enough to be portable.
And carry around and operate continuously.
And we've done this in the form of a device we call a neurochip, so this
neurochip is shown here. Basically it can consists of a printed
circuit board that is connected to electrodes in the brain.
Printed circuit board is populated with off the shelf components including an
amplifier a computer chip and a stimulator.
So the computer chip can be programmed to do a number of operations on the recorded
activity. For example, to identify the occurrence
of action potentials of a single cell at the recording electrode.
And then deliver, convert those action potentials to deliver stimuli back into
the brain. So what's all this good for?
So the bidirectional brain computer interfaces basically have two general
types of applications. one is to, they said create these
artificial recurrent connections. And the fact that the brain can adapt to
consistent sensorimotor conditions means that it could actually learn to
incorporate this bidirectional Brain Computer Interface into normal behavior.
There are obvious clinical applications to using this artificial recurrent
connection to bridge lost biological connections.
So the second general type of application is to create synaptic plasticity.
And this occurs because spike-triggered stimulation produces a synchrony that's
required to strengthen synaptic connections.
Again, a clinical application would be to strengthen weakened connections as occur
in a stroke. So, I'm going to show you two examples
from our lab one of each of these applications.
So first to look at the possibility of creating a recurrent connection that
bridges a lost biological connection. This experiment with Chuck Moritz and
Steve Perlmutter showed that monkeys could learn to use this bidirectional
computer interface. To, control electrical stimulation of
muscles triggered on neural activity in the brain.
So the experiment went as follows, first, the monkey was controlling a cursor, and
driving it into a target by normal, forces generated around the wrist.
The hand was held in an isometric torque transducer, and the monkey generated
these torques, that drove the cursor into the target.
So then, the peripheral nerve was blocked, and this paralyzed the muscles
temporarily. And then the money was rewarded, and then
the cell activity was connected to a controlled cursor.
And the monkey learned in within 10, 20 minutes to drive the cursor into the
target with neural activity. And the final step was to then connect
the activity of the cell to electrical stimulation of the muscle.
And this generated electrically evoked twitches that now generated forces, and
it was the forces that drove the cursor into the target.
So, the monkey was essentially controlling the cursor position with,
torques that were generated by electrical stimulation that was muscle.
That was controlled by neural activity. So, the, an example of, an operation of
this type is shown in this, slide. So, in this situation, there were
actually two cells, one drove stimulation to flexor muscles.
The second row stimulation to extensor muscles.
And the top trace here shows the forces generated through this stimulation of
muscles. As the monkey tried to acquire visual
targets that are represented by these rectangles.
So here, for example, the rectangle queued flexing forces, the monkey
generated the force by increasing the activity of cell one.
And when cell one's activity exceeded the threshold, the flexor muscles were
stimulated in, with intensities proportional to the difference between
firing rate and threshold. And so this produce greater extension
force. And then in order to get into the
extension target the monkey activated the second cell.
So he could generate alternate flexion extension torque's by altnernately
activating these two cells. Another scenario is one in which the
monkey actually controlled bidirectional movements with only one cell.
But, control this by increases and decreases of the cell activity.
Now here we've got again the same, sort of torque trajectories generated, when
the extension and inflection targets are presented.
So in this case the increase in activity above a certain threshold stimulated
flexor muscles. And in order to stimulate the extensors
this monkey had to decrease the activity below a certain level.
So one can actually use the ability of the animal to increase and decrease
neural activity to generate alternating forces.
Now one of the interesting findings in the study it didn't actually turn out to
be necessary to find cells that were normally related to the risk.
Turns out that any motor cortex cell can be volitionally controlled.
And during this period when the monkey learned to control the cell they learned
that the control of any cell, whether it was related to risk or not.
So the importance of that is that It's not necessary to go looking for cells
that are normally related to risk movements and decode their activity.
It's possible to have the subject learn to control this activity.
and the transition to controlling the activity of a cell to controlling muscles
is relatively Straight forward. So the paradigm is illustrated here in
relation to controlling multiple muscles. So [COUGH], one can imagine multiple
cells being directly connected to activate multiple muscles.
And in principle this could work in practice there are challenges.
because electrical stimulation of the peripheral musculature is recruiting
motor neurons in an artificial way. And it's going to get complicated to be
able to control multiple muscles in a synergistic way.
So, an alternate place to stimulate is actually in the spinal cord, as shown
here. Where neural activity can deliver stimuli
in a spinal cord and those spinal stimulation activate circuitry.
That number one ,recruits the motor units in a more natural way.
And number two often activates muscles in synergistic ways so this is probably a
more practical cellular stimulation. Now, the second example I want to show
you is this plasticity that can be produced by spike-triggered stimulation.
[COUGH]. In this example, used the activity of
cells in the primate motor cortex that have direct connections to muscles.
So, in certain so called cortical motor neuron cells, go from the cortex to the
motor neurons. And these synaptic connections then
activate the motor neurons or facilitate the activity of motor neurons.
Which then produce activity in the peripheral muscles.
And these cells can be identified by so called spike-triggered averaging of the
EMG activity. An example from early experiments shown
here, monkey is doing alternate extension inflection movements against the load.
Here we have six extensory muscles that are coactivated with the extension.
A cell is coactivated with these muscles and using the action potential of these
cells to compute spike triggered averages of the EM, rectified EMG activity.
Shows the two of these muscles showed this post spike facilitation.
Post spike increase that is a signature of this mono synaptic connection to motor
neurons of those two muscles. So this is a measure of the strength of
the connection here. So here again is the circuit showing.
The cortical cell monosynaptic connected to motor neuron and that produces this
post spike facilitation. There are also inhibitory circuits that
can be revealed by post spike suppression and are mediated by a disynapTic link
through inhibitory inner neurons. So, this is a way to show what muscles
the, a given single cortical cell facilities and suppresses.
And over here on the left is a summary diagram of the types of cells that were
found in these earlier studies. The main point here is that many of these
cells had divergent connections to multiple target muscles.
So the conditioning of this connection was performed in a study by Yukiyo
Nishimura and Steve Promotor/g. Where the [COUGH] activity of the
cortical cell was recorded, not any cortical cell but cortical monoclonal
cell. So these where neurons that were
demonstrated to have post spike effect in target muscles.
And the neurochip was used to trigger stimulation in the spinal cord of the
target motor neuron at various delays. So we implemented, in other words,
parallel circuit through the neurochip that operated in parallel with the
descending action potentials of the recorded cell.
And, this, produced synchrony at the target site that, could change these
connections. Now, here's, the, demonstration of, the
change in these connections. It was measured by the magnitude of this
post spike facilitation. So here's the spike triggered average
showing post spike facilitation. Spike-triggered average was compiled on
day zero when the monkey was generating standard movements through isometric
forces. Target tracking task and after the post
spike effect was documented the neurochip was connected and operated for about 22
hours. Producing spike-triggered stimulation at
the the spinal site that where this cell terminated.
And then the second day, the post spike effects were measured again with under
the same conditions. And this, post spike effect increased.
The post spike effect magnitude is measured by the so called mean percent
increase over baseline, which was 6 before and about 10 after.
And, this definitely, is, related to the magnitude of the, strength of the cortico
mono [UNKNOWN] connection. This sort of experiment was performed, in
a number of cell muscle pairs, using, different delays between the spike and
the spinal stimulus. And this curve plots the change in this
mean percent increase as a function of the delay between the spike and the
stimulus. And makes the point that increases in the
mean percent increase were obtained in a specific temporal window.
That is to say, for delays between ten and about 25 milliseconds recorded here.
So there was no change for longer intervals.
Which is interestingly is a control for the effect of stimulation alone.
It required spike-triggered stimulation in this interval.
And this is consistent with the temporal window for spike timing dependant
plasticity. Even more interesting, I think, is this
point here, which [COUGH], represents a decrease in the mean percent increase of
this post spike effect. And this was obtained when there was zero
delay between the spike and the stimulus. What this means is that the spinal motor
neurons were electrically activated prior to the arrival of the descending cortical
spinal action potential. And that produced a decrease in the
strength of the connection which is also consistent now with this bidirectional
spike timing dependent plus this integral.
So these two examples illustrate how one can use the bidirectional Brain Computer
Interface to implement artificial connections.
And also create plasticity. And in conclusion there are lots of
applications for this this paradigm. Depending on the source of the signals
the transform of signals and where the stimulation is delivered.
So The sources include neurons either single cortical neurons, multi unit
activity. Or it could be more easily recorded field
potentials could be less invasively recorded electrocardiogram or EMG.
And the side of recording can be pretty much anywhere in the brain.
So this activity can be transformed by delivering stimuli that were directly
proportional to this activity. or through computed function of this
activity one can even imagine all networks being used to convert this
activity to stimulation. And finally the stimulation can be
delivered at multiple targets, muscles we've seen, the spinal cord or cortex.
Or also another interesting target is subcortical reward site, in other words
inner cranial reinforcement sites where the activity that triggers that
simulation gets rewarded and it basically implements a operand conditioning
paradigm that operates during free behavior.
So these are just examples of some of these multiple applications that are
going to provide exciting insights into basic neuroscience mechanisms.
As well as promising clinical applications.
Thank you.
Explore our Catalog
Join for free and get personalized recommendations, updates and offers.
Coursera provides universal access to the world’s best education,
partnering with top universities and organizations to offer courses online.
© 2019 Coursera Inc. All rights reserved.
